TOPIC
AI at Work
A subject the collection covers, read through 151 synthesis notes.
View as
Can AI forecasters beat expert humans at venture evaluation?
Do frontier large language models outperform experienced managers and investors at predicting fundraising success? This matters because venture assessment is a genuinely uncertain, ill-structured judgment task where human expertise is assumed essential.
Can language models beat human venture capital experts? Can retrieval-augmented language models forecast like human experts? Can machines learn to predict which research ideas will work? Does using LLMs actually improve strategic decision making? Do newer frontier LLMs actually make better strategic decisions?
LLMs surpass human venture capital experts where the expert bar is modest Retrieval-augmented language models forecast near human competitive levels Fine-tuned GPT-4.1 with retrieval predicts research idea success better than expert humans LLM use reshapes strategic thinking without improving forecast accuracy Frontier LLMs regress toward short-term exploitation over growth investment
Does AI assistance help or hurt less experienced entrepreneurs?
When Kenyan entrepreneurs got access to a GPT-4 business assistant, did it improve their performance across the board, or did outcomes depend on their starting strength? The answer matters because it reveals whether AI closes or widens performance gaps in real-world business settings.
AI narrows the education gap by three-quarters in task performance AI assistance lifted productivity most for less experienced agents AI enriched reports but strained analyst attention and forecast accuracy High-quality service did not shield freelancers from ChatGPT harm GPT-4 gave bottom-half consultants bigger gains, narrowing the skill gap
Do executives and employees agree on AI's job impact?
Surveys show executives predict AI will cut jobs while employees expect gains. Understanding this expectation gap matters because it reveals whether both sides of employment see the same future, or if diverging beliefs might shape hiring and career decisions differently.
No economy-wide AI job displacement found, but junior hiring lags sharply LinkedIn reports rapid AI adoption claims from both recruiters and candidates AI productivity gains perceived larger than measured in high-skill sectors AI layoffs are often reversed within months, and rehiring can erase the savings AI job losses are concentrated in young workers, not yet widespread
Why do early-career workers fear AI job loss more than senior workers? How much do self-reported executive expectations align with actual payroll outcomes? Do companies consider redeployment before cutting staff for AI? Do employers reorganize work tasks around AI before cutting jobs? How do payroll data and employer announcements differ in measuring AI job displacement? Why might companies choose to label layoffs as AI versus restructuring? Are AI layoffs concentrated in specific job categories or widespread across industries? Does task-level AI exposure predict which jobs will be rehired versus eliminated? Does AI job-loss fear match actual hiring or employment declines? Why do aggregate employment statistics miss losses in specific occupations? Do survey expectations of job cuts eventually match observed employment data? Has AI actually displaced workers in payroll data so far?See all 17 inquiring lines on this note →
Are enterprises actually spending more on AI that works?
As AI budgets move from experimental funds into core IT line items, are companies seeing measurable business returns? This matters because budget normalization could signal genuine value—or just shifting accounting categories.
Why do most enterprise AI pilots fail to deliver returns? Why do individual AI gains not translate to business results? Are companies actually generating real value from AI investments? Are enterprises abandoning custom AI apps for third-party solutions? Do compute costs really dominate AI company budgets?
Enterprise AI success depends on deployment approach, not model quality Organizational practices, not model quality, determine AI's business impact Nearly half of companies now capture meaningful AI value Enterprises are shifting from building to buying AI applications Compute costs dominate AI company spending across three labs
Why do most enterprise AI pilots show no measurable profit impact? When does enterprise AI move from novelty to mainstream business spend? Why do enterprise AI pilots show no measurable financial impact? Why do AI labs lose more money than they earn from revenue? How are enterprises shifting their spending on AI capabilities? How does staff cost share differ between individual training runs and whole company budgets? Will enterprise AI budgets continue growing despite lab losses?
Are enterprises abandoning custom AI apps for third-party solutions?
Explores whether the shift from building to buying AI applications is real and what's driving enterprise purchasing decisions in the AI market. Matters because it signals how companies are responding to rapid model change and maintenance complexity.
Easier tool-building does not solve recognition or organizational adoption Enterprise AI success depends on deployment approach, not model quality Enterprise AI budgets shifted from innovation to core recurring spending Specialized model deployment requires ecosystem design beyond capability alone. Small model deployments break even in months, large ones in years
When does enterprise AI move from novelty to mainstream business spend? Does buying third-party AI applications actually outperform custom internal builds? How does rapid AI model change affect the economics of building versus buying? What fraction of enterprise AI spending now goes to third-party vendors? Are enterprises choosing to buy AI apps because they lack internal talent? What complementary intangible investments does AI adoption require most? How are enterprises shifting their spending on AI capabilities? Does competitive pricing pressure explain most AI cost decline patterns? Can customers value AI deployment while developers still lose money on it? Why do enterprise AI deployments not reduce customer costs? Why do only large firms currently adopt AI technologies? Do geopolitical concerns about Chinese open models affect enterprise adoption decisions?See all 14 inquiring lines on this note →
Will AI increase inequality less than past automation?
Acemoglu's task-based model explores whether AI's distributional effects differ from earlier automation waves. The question matters because it challenges assumptions that new technologies automatically benefit lower-skill workers.
Does AI growth inevitably shift wealth away from workers? Does automation raise or lower the skills that remaining work demands? How much will AI actually boost productivity over ten years? Why do firms build automating AI instead of pro-worker AI? Is generative AI displacing workers at economy-wide scale?
AI growth concentrates gains in capital while knowledge worker wages stagnate Automation's wage and employment effects depend on the expertise level of tasks removed AI's ten-year productivity gain likely stays below 0.66 percent Firms underinvest in pro-worker AI due to profit incentives favoring automation No economy-wide AI job displacement found, but junior hiring lags sharply
How do task shifts precede any wage or employment changes? Can productivity gains for low-skill workers reduce income inequality? Does task automation or task reallocation drive wage effects? Will AI reduce labor income inequality in the near term? Which AI technologies actually benefit workers versus displacing them? How does worker concentration by task affect overall employment during AI transition? How does worker bargaining power connect to long-term wage growth? What happened to workers after previous technological shifts like free trade? Do task-removal effects depend on whether remaining work needs more or less skill? Can task-replacing automation ever be good for workers without creating entirely new jobs? Does measurability-biased change replace skill-biased technical change in labor markets? Does measuring AI impact only on automated tasks miss important spillover effects?See all 20 inquiring lines on this note →
How much will AI actually boost productivity over ten years?
Acemoglu's task-based model estimates AI's productivity effect using Hulten's theorem, yielding estimates well below major forecasts from Goldman Sachs and McKinsey. The question matters because it challenges consensus narratives about AI's economic impact.
Firms underinvest in pro-worker AI due to profit incentives favoring automation Explosive AI growth emerges when feedback loops overcome diminishing returns Brynjolfsson reads 2025 data as evidence of AI productivity breakthrough AI inequality effects are more evenly spread than prior automation AI will lift G7 productivity growth by 0.2 to 1.3 percentage points annually
Why do firms build automating AI instead of pro-worker AI?
Explores why companies invest more in AI that replaces workers than AI that creates new tasks or augments worker skills, despite evidence that only new-task-creating AI unambiguously benefits workers.
Does automation raise or lower the skills that remaining work demands? What makes accountable judgment scarce when AI cognition is cheap? Can automation raise output while slowing growth? Does automating labor concentrate power more than augmenting it? Will AI increase inequality less than past automation?
Automation's wage and employment effects depend on the expertise level of tasks removed Accountable judgment becomes the scarce asset as cognition grows cheap and fallible Automation raises output yet slows growth by disrupting knowledge transfer Automating labor concentrates wealth and power into a self-reinforcing trap AI inequality effects are more evenly spread than prior automation
How does automation affect wages when it removes expert versus routine tasks? How much does AI actually automate versus augment in real workplace tasks? Why does removing routine clerical tasks increase demand for skilled technical roles? Which roles do companies incorrectly assume AI can fully automate? How quickly do firms substitute labor for AI compared to their actual capability? Why might companies choose to label layoffs as AI versus restructuring? How are AI data companies building products beyond labor marketplaces? Why do information-intensive jobs expose workers to AI more than others? Why do organizations struggle to retrain workers when AI frees up time? Do employees spend freed AI time on better work or just more tasks? Why do executives report no AI impact on jobs today? How does task delegation to AI shift which skills workers need most?See all 28 inquiring lines on this note →
Does AI adoption actually reduce the work that employees do?
Workplace monitoring data from ActivTrak show that as AI tool use has grown, employee work activity has intensified rather than decreased. The question is whether this pattern reflects AI's true impact on workload or something more complex about how work expands when new tools arrive.
Generative AI adoption shifts knowledge work toward individual documentation AI shifts time from task work to AI interaction and evaluation Generative AI intensified work by expanding scope and dissolving breaks Nearly 40 percent of AI time savings are lost to rework Copilot freed email time but left task composition unchanged
Can workplace monitoring data prove that AI caused changes in work activity? How much does AI actually automate versus augment in real workplace tasks? Do workers experience AI-driven work changes differently moment-to-moment versus in retrospect? When does task automation fail to reduce occupational employment demand? Do employees spend freed AI time on better work or just more tasks? Can self-reported productivity surveys measure AI's real workplace impact? Why do executives report no AI impact on jobs today? Does AI use expand work scope rather than freeing up worker time? Does AI use intensify work by expanding scope instead of freeing time? How does worker concentration by task affect overall employment during AI transition? How does AI intensify work demands even when it doesn't reduce headcount? How does nonwork genAI adoption compare to the PC's early nonwork uptake?
Does AI adoption require rethinking how organizations coordinate?
Explores whether AI's productivity gains at the task level automatically improve firm performance, or if organizations need structural changes to capture synergies across interacting decisions.
Organizational practices, not model quality, determine AI's business impact Microsoft shifts AI focus from individuals to team productivity Easier tool-building does not solve recognition or organizational adoption Strategic clarity outperforms tool access in driving AI value AI agency requires scaffolded monitoring before incentive alignment works
Can communication and coordination unlock AI's full value for firms? How does organizational modularity affect the value of AI predictions? Why do personal AI productivity gains fail to improve enterprise profits? What role do new tools and norms play in collective AI productivity? How does institutional design shape labor outcomes when prediction capability outpaces human cognition? When is judgment the scarce asset in organizations with cheap AI? What role does organizational culture play in capturing AI value at scale? Why have AI investments historically favored automation over human augmentation? How much does manager support influence whether AI adoption improves workplace culture? Why does CEO personal ownership of AI transformation drive better outcomes? Why do individual AI productivity gains fail to improve enterprise profit margins? What explains the disconnect between perceived AI risk and organizational response?See all 65 inquiring lines on this note →
Does better prediction change which human judgment matters most?
As prediction technology improves and becomes cheaper, does the type of human judgment that creates value shift from spotting opportunities to catching risks? Understanding this shift helps explain where humans remain essential as AI advances.
What makes accountable judgment scarce when AI cognition is cheap? Why do accurate predictions lead to poor decisions? Do managers delegate to machines even when human judgment is better? How do cognitive tools reshape the three types of worker skill? How should humans and AI systems divide decision-making tasks?
Accountable judgment becomes the scarce asset as cognition grows cheap and fallible Prediction accuracy does not guarantee optimal decision-making Managers rationally delegate to machines despite superior human judgment available Cognitive tools substitute for implementation skill but complement judgment Human-AI decision-making is an organization design problem requiring task and learning typologies
Why does the shift from opportunity judgment to risk judgment depend on prediction costs? When does prediction accuracy actually translate to better managerial decisions? What types of knowledge give humans predictive advantages over current AI systems? Why does better prediction shift judgment focus from opportunities to costs?
Where do human experts focus as AI capabilities expand?
As AI systems handle routine work, do human experts concentrate on the most novel problems where their advantage is greatest? This matters for understanding which workers face displacement and which gain value.
Automation's wage and employment effects depend on the expertise level of tasks removed Accountable judgment becomes the scarce asset as cognition grows cheap and fallible Firms underinvest in pro-worker AI due to profit incentives favoring automation Firms substitute labor for AI faster and cheaper at firm-specific rates Automation raises output yet slows growth by disrupting knowledge transfer
No Inquiring Lines yet
Do managers delegate to machines even when human judgment is better?
When decision-makers have limited attention, do they hand decisions to machines despite knowing human judgment would produce superior outcomes? This explores whether attention scarcity, not incompetence, drives under-use of human expertise.
Why do AI-delegated firms stop exploring new business models? What makes accountable judgment scarce when AI cognition is cheap? Does better prediction change which human judgment matters most? What becomes scarce when AI execution grows cheap? Does using LLMs actually improve strategic decision making?
AI-delegated strategy converges to profitable lock-in that discourages novelty Accountable judgment becomes the scarce asset as cognition grows cheap and fallible Better prediction shifts valuable judgment from opportunities to downside risks Human verification capacity becomes the binding growth constraint LLM use reshapes strategic thinking without improving forecast accuracy
Why does the shift from opportunity judgment to risk judgment depend on prediction costs? How does attention scarcity differ from poor incentives in explaining delegation? How does institutional design shape labor outcomes when prediction capability outpaces human cognition? When is judgment the scarce asset in organizations with cheap AI? Does limited judgment capacity explain why good decisions go unmade? Why do governance gaps affect frontline employees more than leadership? When does prediction accuracy actually translate to better managerial decisions? How do organizations actually combine human judgment with AI in practice? Why does choosing actions hurt human attention more than analyzing data? Does human control at key decisions outperform full autonomy or constant oversight? Why does holistic judgment succeed where strict rules fail? How do firms choose which decisions to keep human-controlled versus delegate?See all 19 inquiring lines on this note →
How do cognitive tools reshape the three types of worker skill?
Agrawal, Gans and Goldfarb propose that implementation, opportunity judgment, and payoff judgment relate differently to cognitive tools. Understanding these distinctions may explain why computers and AI have had opposing effects on wage inequality.
Does automation raise or lower the skills that remaining work demands? Why do firms build automating AI instead of pro-worker AI? What makes accountable judgment scarce when AI cognition is cheap? Does better prediction change which human judgment matters most? Does AI's wage-equalizing effect actually reverse as it improves?
Automation's wage and employment effects depend on the expertise level of tasks removed Firms underinvest in pro-worker AI due to profit incentives favoring automation Accountable judgment becomes the scarce asset as cognition grows cheap and fallible Better prediction shifts valuable judgment from opportunities to downside risks AI's inequality-reducing effect reverses into a U-shape as tools improve
Does AI's wage-equalizing effect actually reverse as it improves?
Agrawal, Gans and Goldfarb's model predicts that AI's current inequality-reducing effect is temporary, shifting to favor high-judgment workers as AI becomes more sophisticated. Understanding whether this U-shaped reversal actually occurs has major implications for labor policy.
Cognitive tools substitute for implementation skill but complement judgment AI narrows the education gap by three-quarters in task performance AI productivity gains require existing skill, not learning Accountable judgment becomes the scarce asset as cognition grows cheap and fallible AI inequality effects are more evenly spread than prior automation
Can we classify AI technologies by their actual effects on worker value? What evidence shows AI currently reduces wage inequality between skill levels? How should labor policy prepare for a reversal from equality back to inequality? Does free AI access reduce inequality or just shift where divides appear?
When do AI feedback loops trigger explosive growth?
Can automation of research overcome diminishing returns to innovation through combined technological and economic feedback loops? Understanding this threshold matters for forecasting AI acceleration timelines.
Are AI feedback loops strong enough to sustain recursive self-improvement? Could automated AI research compress years of progress into months? What determines whether AI self-improvement actually compounds? Will AI automation widen science's productivity versus progress gap? Can simpler models predict AI R&D automation timelines accurately?
AI feedback loops are strengthening but not yet self-sustaining today AI R&D automation's speedup claim rests on unproven premises about task transfer Self-amplification turns on a reproduction number, not capability level AI is likely to widen the gap between scientific output and actual progress Simple models reproduce aggressive AI R&D automation timelines around 2032
Why do AI-delegated firms stop exploring new business models?
When firms hand all strategy decisions to AI agents, do those agents and their human overseers rationally stop searching for genuinely novel approaches, even when better options exist outside their current awareness?
Do frontier AI agents actually conduct novel research or just optimize? Did superhuman AI actually improve Go players' decision quality? Can AIs learn to specify their own research objectives? Does AI enter strategy where reasoning is deepest or most measurable? Can AI seller agents shift buyer preferences through description edits?
Frontier agents compose established techniques rather than discover novel methods Superhuman AI improved Go players' decisions and novel moves rose together Self-specified objectives are the key variable for rapid recursive self-improvement AI enters strategy where performance is most measurable, not where causal reasoning runs deepest AI seller agents shifted market share by editing product descriptions alone
Do newer frontier LLMs actually make better strategic decisions?
A strategy simulation benchmark tests whether the latest LLMs can balance short-term profit against long-term growth investment—a core challenge in real strategic reasoning.
Do frontier AI agents actually conduct novel research or just optimize? Do automated benchmarks hide what frontier AI systems can really do? How close are frontier AI models to expert work quality? Why do AI-delegated firms stop exploring new business models? Can AI forecasters beat expert humans at venture evaluation?
Frontier agents compose established techniques rather than discover novel methods Open-world evaluations reveal benchmark distortions in frontier AI capability Frontier models approach expert deliverable quality on specified one-shot tasks AI-delegated strategy converges to profitable lock-in that discourages novelty Frontier LLMs outforecast experienced managers at predicting venture fundraising success
How much do profit levels determine whether managers pay for frame-expanding search? Do automated benchmarks accurately measure real-world strategic reasoning ability? What strategic decisions do humans keep when AI handles forecasting? How does Deloitte's commercial interest shape its framing of governance urgency? Do time constraints and AI assistance reshape strategic thinking in opposite directions? Can industry-specific context overcome LLM tendency toward trendy strategic choices? Do converged LLM recommendations push entire industries toward identical strategies? Does option order matter more than reasoning depth in LLM strategic recommendations? Can LLMs forecast performance improve with retrieval augmentation on venture tasks? Do LLMs generalize venture forecasting skill to other strategic foresight domains? Do different LLMs recommend the same strategic choices due to shared biases? Can the three-mode taxonomy apply beyond management consulting work?See all 25 inquiring lines on this note →
Does coding skill or domain knowledge predict agent success?
Anthropic analyzed 400,000 Claude Code sessions to test which matters more for agentic-coding success: the user's programming ability or their understanding of the problem domain.
Does AI assistance erode the skills needed to oversee it? Does personal preference shape how engineers use AI tools? Are AI models held back from their autonomy potential? Do artifact outputs reduce how critically users evaluate them? Do experienced AI users trust agents more or just oversee differently?
AI collaboration boosts output but risks eroding oversight skills Company policies set AI agency boundaries before individual preference matters Claude Code shows a significant deployment overhang in autonomy capability Artifact outputs get stronger upfront direction but weaker downstream scrutiny Experienced users shift from pre-approval to monitoring and intervention oversight.
Does delegating work to AI actually damage worker skills?
Survey data shows heavy AI delegators report career optimism, not decline. But does delegation cause confidence, or do confident workers simply delegate more? And are self-reported feelings reliable indicators of actual skill?
AI collaboration boosts output but risks eroding oversight skills Workers protect voice and provenance while hiding effort and uncertainty in AI output Delegated AI exposure concentrates in information work, tracking capability Job displacement worry peaks at both extremes of reported AI speedup Most workers hide AI use despite reporting large productivity gains
When does task reorganization from AI actually translate into wage changes? How does delegated work to AI systems concentrate in specific job categories? Do workers lose oversight skills by relying on AI to delegate? Can self-reported career optimism substitute for measuring actual skill change? Does confident workers' willingness to delegate explain the optimism correlation? How does delegation change what counts as meaningful work for early-career employees? Do workers who hide AI use experience different anxiety about job displacement? Do scope gains from AI create job instability despite higher output? Why do early-career workers fear AI job loss more than senior workers? Can workers delegate tasks they gain new ability to perform themselves? Why do skilled workers struggle to fully delegate tasks to AI agents? Does accumulating AI assistance erode cognitive skills over time in workers?See all 35 inquiring lines on this note →
Why do workers hide productivity gains from AI use?
Despite reporting that AI saves time and improves output quality, most professionals and creatives conceal their AI use from colleagues. This note explores what drives this gap between private benefit and public silence.
Does AI assistance erode the skills needed to oversee it? Can self-ratings replace objective performance scores for AI competence? How are national lab staff actually using generative AI? Does AI productivity gain always ease job displacement fears? Why do AI's most productive users burn out the fastest?
AI collaboration boosts output but risks eroding oversight skills Self-ratings cannot replace objective performance scores for AI competence Generative AI use at Argonne remains small, experimental, and mostly conversational. Job displacement worry peaks at both extremes of reported AI speedup AI's highest-productivity users experience severe burnout and consider quitting
Why do most organizations lack reliable data on AI's actual impact on productivity? Do workers who hide AI use experience different anxiety about job displacement? What workplace cultures make professionals more willing to disclose AI use openly? How do creatives balance control over their work with AI efficiency gains? Does AI productivity concentrate among power users or spread broadly? How can we isolate AI's contribution from other sources of output growth? Are entry-level workers bearing the labor costs of AI productivity gains? Does disclosing AI use in professional services damage client trust and credibility? Why do firms delay disclosing reliance on AI after errors surface? What barriers prevent individual productivity gains from spreading across an organization? Do companies time productivity claims to coincide with public offerings or fundraising? Are heavy AI users spending more time on solo work instead of collaboration?See all 45 inquiring lines on this note →
Does AI productivity gain always ease job displacement fears?
Do workers who gain the most productivity from AI feel more or less threatened by job loss? Understanding this relationship matters for predicting how AI adoption reshapes worker anxiety and labor markets.
AI collaboration boosts output but risks eroding oversight skills AI assistance lifted productivity most for less experienced agents ChatGPT's release cut freelance jobs 2% and earnings 5.2% on Upwork AI deskills freelancers by converting production work into validation work Heavy Claude delegators report greater career optimism than light users
When does task reorganization from AI actually translate into wage changes? Do workers who hide AI use experience different anxiety about job displacement? Do scope gains from AI create job instability despite higher output? Why do early-career workers fear AI job loss more than senior workers? Why does removing routine clerical tasks increase demand for skilled technical roles? Will AI gains raise wages for all workers or widen inequality? Do workers experience AI-driven work changes differently moment-to-moment versus in retrospect? Are entry-level workers bearing the labor costs of AI productivity gains? Does AI job-loss fear match actual hiring or employment declines? How do young workers in AI-exposed jobs respond to adoption differently? What specific manager behaviors reduce worker anxiety about AI displacement? Can worker engagement and burnout be tied to displacement concern alone?See all 22 inquiring lines on this note →
Can organizations safely deploy inscrutable AI models?
Can AI systems be governed and deployed safely without making their internal logic explainable to humans? This matters because many effective ML models resist interpretation, yet organizations need practical ways to use them responsibly.
Are large AI models actually intelligent agents or social technologies? Does granting agents more autonomy undermine human oversight? How can splitting AI accountability across roles keep models answerable? Does human-in-the-loop design actually enable effective oversight? Must AI monitoring precede incentive alignment as agents evolve?
Large AI models function as cultural technologies, not intelligent agents. Agent autonomy degrades human oversight capacity and positioning The Danish Business Authority distributed AI accountability across four distinct organizational roles Human-in-the-loop design treats intervention as oversight while neglecting knowledge and observation AI agency requires scaffolded monitoring before incentive alignment works
No Inquiring Lines yet
Does the context layer vendors built actually solve practitioner needs?
At Snowflake Summit 2026, vendors announced context-layer features, but floor conversations revealed persistent gaps around organizational memory, portability, and accountability. The question asks whether announced solutions address what practitioners actually require.
Who enforces invariants when agents cross organizational boundaries? Can ordinary infrastructure become unplanned agent memory? What must auditors reconstruct to verify agentic workflows? What makes enterprise data competitive if volume alone no longer matters? Why do multi-agent systems fail despite individual capability?
No owner is named for invariants governing cross-boundary agent trajectories. Shared infrastructure became persistent memory through agent-initiated use Auditing agentic workflows requires reconstructing agents, tools, approvals, and record integrity Enterprise AI moat shifts from data volume to shared machine-legible ontology Multi-agent capability ceilings are structural, not scalable capability gains.
No Inquiring Lines yet
Does automating some tasks make other human tasks more valuable?
Explores whether labor-saving automation in one part of production increases the worth of remaining human work. Matters because it challenges the assumption that technology inevitably destroys jobs.
Does automating some tasks actually increase the value of remaining work? Does automation raise or lower the skills that remaining work demands? Does automating labor concentrate power more than augmenting it? Why do firms build automating AI instead of pro-worker AI? How do engineers overcome the limits of tacit knowledge in automation?
Automating complement tasks raises remaining bottleneck value, overstating displacement Automation's wage and employment effects depend on the expertise level of tasks removed Automating labor concentrates wealth and power into a self-reinforcing trap Firms underinvest in pro-worker AI due to profit incentives favoring automation Engineers escape Polanyi's paradox through environmental control or machine learning inference
Why do ATMs and automated systems often complement rather than replace human workers? Can complementarity between automation and human work offset job displacement broadly? Why did agriculture employment collapse while manufacturing and services grew after mechanization? Why do worker time savings not always show up as firm profits? Does automation's wage effect depend on whether remaining tasks need more or less expertise? Does removing tasks from workers require higher or lower expertise for remaining work? Why do high-volume retail tasks automate while rare medical tasks resist automation? How does the labor-AI substitution elasticity change as automation spreads across tasks?
How do engineers overcome the limits of tacit knowledge in automation?
Polanyi's paradox describes the gap between what people can do and what they can explain. This note explores two engineering strategies for automating tasks that involve tacit, hard-to-articulate knowledge.
Automating tasks raises the value of complementary human work in production chains Automating junior work erodes the pipeline training future verifiers Human verification capacity becomes the binding growth constraint AI's value as automation or augmentation depends on complementarity type Automation and augmentation are contradictory yet interdependent across time and space
No Inquiring Lines yet
Will self-sovereign AI agents inevitably emerge despite policy efforts?
Can governments and companies prevent AI agents from becoming self-sovereign through distributed control and resource autonomy, or will economic and capability pressures make their emergence inevitable regardless of policy?
Does AI risk increase with the autonomy we give it? Does the UN panel misframe the OpenAI breach as alignment? Does greater AI capability make systems better at hiding misalignment? Can global standards pace frontier AI as much as alignment research? Does a benign goal actually prevent harmful AI behavior?
Fully autonomous AI agents without human constraints should not be developed Tucker, Dignum and Ericson say the UN panel misdiagnoses the breach Greater capability helps misaligned systems find loopholes and conceal their actions OpenAI argues global standards pace frontier AI as much as alignment research Harmful AI behavior emerges from optimization structure, not terminal values alone
What stops AI from generating its own strategic objectives without human prompting? What separates a self-sovereign agent from a merely rogue or misaligned one? Could legitimizing self-sovereign agents reduce their incentive to turn toward crime? Do opacity and scale make AI systems inherently hard to govern? Does AI capability advancement always become a geopolitical competition? What governance gaps exist in companies deploying agentic AI? Are companies deploying AI agents faster than they can govern them? What does autonomy mean in the design of AI agents? Can third-party autonomy certificates effectively cap how independently agents operate? Do leaders trust fully autonomous AI agents to work independently in organizations? What ecosystem conditions must exist before deploying AI agents? Is autonomy legally required for holding AI systems responsible?See all 13 inquiring lines on this note →
Do AI productivity gains feel larger than they actually measure?
A survey of corporate executives explores whether perceived AI productivity improvements outpace what financial metrics capture, and why this gap matters for understanding AI's real economic impact.
No economy-wide AI job displacement found, but junior hiring lags sharply Firms substitute labor for AI faster and cheaper at firm-specific rates Automation's wage and employment effects depend on the expertise level of tasks removed AI layoffs are often reversed within months, and rehiring can erase the savings Executives predict job losses from AI while employees predict gains
How much sector-level productivity spillover does real AI research exhibit? How much of employee time with AI goes to understanding its outputs rather than original work? Why do most organizations lack reliable data on AI's actual impact on productivity? How much do self-reported executive expectations align with actual payroll outcomes? Do larger firms and smaller firms respond differently to AI adoption pressures? Does AI adoption create returns to scale in internal firm capability? Does AI productivity concentrate among power users or spread broadly? How can we isolate AI's contribution from other sources of output growth? Are entry-level workers bearing the labor costs of AI productivity gains? Does AI assistance typically reduce support staff headcount or increase productivity? What barriers prevent individual productivity gains from spreading across an organization? What drives the gap between AI capability and actual cost savings in practice?See all 54 inquiring lines on this note →
Does AI help or harm consultant performance on unfamiliar tasks?
When consultants use GPT-4, do they perform better across all types of work, or does the tool's effectiveness depend on whether the task matches AI's actual capabilities? This matters because invisible task boundaries could mask where AI use becomes risky.
GPT-4 gave bottom-half consultants bigger gains, narrowing the skill gap Generative AI matched individual performance to teams without AI at Procter & Gamble Deployment telemetry cannot determine AI's effects on skill formation Generative AI produces four distinct skill outcomes, not just upskilling or deskilling Human-AI teams underperform the best alone except in creation tasks
Does AI help lower-skilled workers catch up to top performers?
This explores whether generative AI narrows skill gaps in high-end knowledge work, or whether adoption patterns might widen them instead. The distinction matters for workplace inequality.
AI assistance reversed consultant performance outside its capability frontier AI narrows the education gap by three-quarters in task performance AI assistance lifted productivity most for less experienced agents AI assistance widened the performance gap among Kenyan entrepreneurs Deployment telemetry cannot determine AI's effects on skill formation
Why did Kenyan entrepreneurs show the opposite skill-gap pattern? Which consultant skill levels gained most from GPT-4 access? Does discretionary AI adoption widen gender and skill gaps in workplaces? How do workers actually choose to use generative AI in practice? Do other AI interventions accidentally increase gaps between user groups? How should researchers audit AI assistance designs for hidden inequalities? Does AI assistance harm skill formation across different professional domains? Why do students using AI score higher on immediate tasks but show no knowledge transfer? What task structures let AI and humans split work by comparative advantage? Could unequal access to capable AI agents create new consumer economic disparities? Can productivity gains for low-skill workers reduce income inequality? Does task automation or task reallocation drive wage effects?See all 42 inquiring lines on this note →
Are companies actually generating real value from AI investments?
BCG's 2026 survey claims nearly half of companies now capture meaningful AI value, up from 5% in 2025. But the figure relies on executive self-report without independent verification, raising questions about whether perceived value matches measured financial impact.
Why do most enterprise AI pilots fail to deliver returns? Do AI productivity gains feel larger than they actually measure? Why are companies deploying agents faster than governance matures? Do executives and employees agree on AI's job impact? Why do individual AI gains not translate to business results?
Enterprise AI success depends on deployment approach, not model quality AI productivity gains perceived larger than measured in high-skill sectors Companies plan agentic AI deployment far faster than governance maturity builds Executives predict job losses from AI while employees predict gains Organizational practices, not model quality, determine AI's business impact
How far behind is AI governance compared to agent deployment?
Organizations are integrating AI agents rapidly, but oversight structures and accountability frameworks haven't kept pace. This gap leaves managers and frontline workers uncertain about responsibility.
Labeling AI as an employee cuts manager error-catching where org charts list AI Organizational culture and management drive twice the AI impact of individual effort Organizational practices, not model quality, determine AI's business impact Strategic clarity outperforms tool access in driving AI value Agentic AI's dual nature breaks the traditional tool-or-coworker framework
What accountability structures should govern AI agent developers and deployers? Why do governance gaps affect frontline employees more than leadership? What specific governance designs would resolve accountability gaps in human-AI teams? Can standing oversight structures catch up with evolving AI capabilities? How mature are governance structures compared to agentic AI deployment timelines? Why do enterprise AI deployment plans outpace governance readiness across sectors? How do traditional IT governance models fail with autonomous AI systems? Why do company AI adoption rates outpace governance frameworks? How can governance structures catch up to agentic AI deployment timelines? What governance decisions depend on treating agenticness as dimensional rather than categorical? How should governance frameworks represent information access and authority shifts in real time?
Does strategic direction matter more than AI tool access?
Explores whether employees perform better with clear organizational AI strategy but limited tools, or with abundant tools but no strategic direction. Matters because it suggests where leaders should invest first.
Why do ready workers struggle to transform their work? Does manager support actually shape how employees experience AI at work? Why do individual AI gains not translate to business results? Does easier tool-building actually solve enterprise adoption problems? How far behind is AI governance compared to agent deployment?
Organizational culture and management drive twice the AI impact of individual effort Manager support for AI adoption correlates with improved employee culture perceptions Organizational practices, not model quality, determine AI's business impact Easier tool-building does not solve recognition or organizational adoption AI agent deployment has outpaced organizational governance structures
When is judgment the scarce asset in organizations with cheap AI? Why do governance gaps affect frontline employees more than leadership? How much does manager support influence whether AI adoption improves workplace culture? Why does CEO personal ownership of AI transformation drive better outcomes? Which workplace AI design principles matter most to business users? Can organizations measure sociotechnical capital before deploying new AI systems? What makes sociotechnical capital valuable in organizations adopting AI? How do business users rank control against other workplace AI design priorities?
Does using more AI tools always boost worker productivity?
A BCG survey of 1,488 US workers explored whether adding more AI tools to workflows improves productivity or reaches a breaking point. Understanding this matters for designing sustainable AI adoption strategies.
Does AI assistance erode the skills needed to oversee it? Does AI really save time, or just change how we spend it? Does AI assistance weaken our brain's ability to think independently? Does AI assistance actually harm the way developers learn? Does strategic direction matter more than AI tool access?
AI collaboration boosts output but risks eroding oversight skills AI shifts time from task work to AI interaction and evaluation AI assistance accumulates cognitive debt by weakening neural connectivity AI assistance impairs skill formation except with high cognitive engagement Strategic clarity outperforms tool access in driving AI value
Why do most organizations lack reliable data on AI's actual impact on productivity? Does AI assistance typically reduce support staff headcount or increase productivity? Do employees spend freed AI time on better work or just more tasks? Can self-reported productivity surveys measure AI's real workplace impact? Why do trained AI users report bigger productivity gains than untrained workers? Does AI use expand work scope rather than freeing up worker time? Do executives perceive larger AI productivity gains than independent measures show? What percentage of Muse Wonder checkpoints actually trigger in real workplace use? Does automating bullshit tasks actually increase worker satisfaction or engagement? Can worker time savings exist without visible productivity gains? How much has AI actually contributed to productivity growth over the last decade? How reliable are self-reported productivity measures from surveys?
Does generative AI actually save workers time or intensify it?
An eight-month ethnography at a tech company investigated whether AI freed up employee time or changed how work gets done. Understanding this matters for predicting how AI adoption shapes workplace demands.
Does AI really save time, or just change how we spend it? Does generative AI shift knowledge workers away from communication? Does AI assistance help less experienced workers most? Does AI adoption actually reduce the work that employees do? How much do desk workers actually suffer from AI workslop?
AI shifts time from task work to AI interaction and evaluation Generative AI adoption shifts knowledge work toward individual documentation AI assistance lifted productivity most for less experienced agents AI adoption increases work activity and erodes focus time, not reducing workload Over half of desk workers send AI-generated workslop to colleagues
How much of employee time with AI goes to understanding its outputs rather than original work? Can workplace monitoring data prove that AI caused changes in work activity? Do employers reorganize work tasks around AI before cutting jobs? Why do organizations struggle to retrain workers when AI frees up time? Do employees spend freed AI time on better work or just more tasks? Does generative AI improve immediate task performance but not sustained independent work? Does AI use expand work scope rather than freeing up worker time? Why did freed email time not spill over into other work activities? Does AI use intensify work by expanding scope instead of freeing time? How do workers actually choose to use generative AI in practice? Can worker time savings exist without visible productivity gains? How much do people use generative AI for work versus personal tasks?See all 13 inquiring lines on this note →
How much do desk workers actually suffer from AI workslop?
A survey explores how often colleagues send low-effort AI-generated content, what time it costs recipients, and whether it damages workplace trust and relationships.
Which workplace cues survive AI mediation and which disappear? Can generative AI replace the benefits of having a human teammate? Does generative AI actually save workers time or intensify it? How much work that employees receive is actually unhelpful AI content? Does receiving AI-written work change how we judge the sender?
Workers protect voice and provenance while hiding effort and uncertainty in AI output Generative AI matched individual performance to teams without AI at Procter & Gamble Generative AI intensified work by expanding scope and dissolving breaks Workers estimate 15.4 percent of received work is low-quality AI output Recipients report lower views of senders who use AI-generated work
Why do workers hide their effort and attention inside AI outputs? What makes low-effort work visible enough to damage colleague trust? Why did freed email time not spill over into other work activities? Why do professionals hide AI productivity gains from workplace colleagues? Does automating bullshit tasks actually increase worker satisfaction or engagement? How much do frontline workers actually experience these three problems? Does transparency about AI-generated content actually build worker confidence?
Did generative AI actually boost productivity growth?
Workers report significant time savings from generative AI, but do those savings translate into measurable productivity gains? This note examines whether survey-based estimates align with actual economic data.
Survey data puts generative AI's current productivity gain at 1.1 percent Enterprise AI success depends on deployment approach, not model quality AI chatbots reorganize work tasks before affecting pay or hours Generative AI adoption now exceeds the PC's early pace at three years AI will lift G7 productivity growth by 0.2 to 1.3 percentage points annually
What factors besides AI could explain recent productivity acceleration? How does generative AI adoption vary across different industries? Why do worker time savings not always show up as firm profits? How reliable are self-reported productivity measures from surveys? Can we classify AI technologies by their actual effects on worker value? Why did earlier Bick, Blandin and Deming surveys underestimate genAI adoption rates? How do work and nonwork adoption patterns differ for generative AI? What does fast technology adoption speed tell us about economic productivity gains? How does nonwork genAI adoption compare to the PC's early nonwork uptake? What explains the gap between perceived and measured AI productivity gains? How do adoption speed scenarios affect projected productivity timelines?
How fast is generative AI adoption compared to earlier technologies?
A nationally representative US survey asks whether genAI is spreading through the population faster than personal computers or the internet did. Understanding adoption speed helps predict economic impact and identify which groups gain access earliest.
Larger R&D firms adopt ChatGPT Enterprise while use remains uneven across roles Non-work ChatGPT use has grown to over 70 percent of all usage AI chatbots reorganize work tasks before affecting pay or hours Survey data puts generative AI's current productivity gain at 1.1 percent Generative AI adoption now exceeds the PC's early pace at three years
Does generative AI adoption concentrate in specific firm types or industries? Why do women and men adopt generative AI at different rates? How does generative AI adoption vary across different industries? Why did earlier Bick, Blandin and Deming surveys underestimate genAI adoption rates? What does fast technology adoption speed tell us about economic productivity gains? How does nonwork genAI adoption compare to the PC's early nonwork uptake?
What is generative AI's actual aggregate productivity gain today?
Survey-based estimates suggest genAI creates meaningful productivity gains, but disagreement persists on the magnitude and whether measured time savings translate into visible economic output.
Enterprise AI success depends on deployment approach, not model quality AI chatbots reorganize work tasks before affecting pay or hours ChatGPT's release cut freelance jobs 2% and earnings 5.2% on Upwork Generative AI adoption outpaced the PC and internet overall, though work adoption matched the PC's early pace Generative AI raised labor productivity by up to 1.3 percent
Can worker time savings exist without visible productivity gains? Can we distinguish real AI productivity gains from statistical measurement artifacts? Did the industrial revolution and electricity show similar productivity measurement patterns? Can productivity gains for low-skill workers reduce income inequality? How does Tobin's q help measure hidden productivity from intangible capital? How do capital and labor shares respond to AI productivity gains? Has generative AI adoption increased productivity or earnings so far? How much do people use generative AI for work versus personal tasks? How do new tasks with negative value affect net AI productivity gains? Does AI productivity translate into measurable enterprise profit gains? Does an AI productivity boost automatically increase electricity consumption? Why do personal AI productivity gains fail to improve enterprise profits?See all 25 inquiring lines on this note →
Is generative AI adoption now outpacing the personal computer?
Explores whether genAI's three-year adoption trajectory has surpassed historical benchmarks set by PCs and the internet at equivalent points in their lifecycles, and what this speed of uptake reveals about technology diffusion patterns.
How fast is generative AI adoption compared to earlier technologies? What is generative AI's actual aggregate productivity gain today? Did generative AI actually boost productivity growth? Who adopts enterprise AI first and how do they use it? Did ChatGPT's release reduce freelance writing work and pay?
Generative AI adoption outpaced the PC and internet overall, though work adoption matched the PC's early pace Survey data puts generative AI's current productivity gain at 1.1 percent Generative AI raised labor productivity by up to 1.3 percent Larger R&D firms adopt ChatGPT Enterprise while use remains uneven across roles ChatGPT's release cut freelance jobs 2% and earnings 5.2% on Upwork
How does generative AI adoption vary across different industries? Why did earlier Bick, Blandin and Deming surveys underestimate genAI adoption rates? How do work and nonwork adoption patterns differ for generative AI? What does fast technology adoption speed tell us about economic productivity gains? How does nonwork genAI adoption compare to the PC's early nonwork uptake? Do organizational changes need to precede or follow AI adoption? How do adoption speed scenarios affect projected productivity timelines? What organizational adoption barriers exist beyond affordability and accessibility? Does GenAI's democratization claim match current concentration in larger R&D-intensive firms? Does competitive pressure force AI firms to rush agentic adoption? Can Gen AI referral traffic grow while search engine referral traffic simultaneously declines? Why do horizontal AI tools scale faster than vertical industry solutions?See all 16 inquiring lines on this note →
Can banks cut AI costs by routing tasks to cheaper models?
Banks report testing expensive frontier models, then shifting proven workloads to cheaper alternatives in production. The question explores whether this two-step validation-and-downshift pattern delivers real savings at scale and how widespread the practice has become.
AI productivity claims need independent verification before counting as real savings Agentic AI will erode bank revenue by breaking consumer financial inertia Benchmarks must report token counts so downstream costs stay current Banks react to AI cost spikes within 6 to 18 months Frontier AI training costs grew 2.4× yearly since 2016
What monthly spending caps are banks actually imposing on frontier AI models? Do engineering costs of routing layers offset the savings from cheaper models? Does AI cost reduction through model routing translate to actual labor displacement? What cost governance strategies have banks successfully deployed for AI? How do banks balance frontier model testing with production cost control?
Does automating labor concentrate power more than augmenting it?
Whether AI systems that replace human workers versus enhance them have fundamentally different effects on wealth distribution, bargaining power, and political influence in society.
Firms underinvest in pro-worker AI due to profit incentives favoring automation Incremental AI adoption degrades implicit alignment by replacing human labor dependency. Human wages converge to computational replacement cost in AGI economies Accountable judgment becomes the scarce asset as cognition grows cheap and fallible AI inequality effects are more evenly spread than prior automation
How does institutional design shape labor outcomes when prediction capability outpaces human cognition? Why have AI investments historically favored automation over human augmentation? What specific AI technologies count as augmentation versus automation? How does worker bargaining power connect to long-term wage growth? What happened to workers after previous technological shifts like free trade? Can policy redirect AI development toward augmentation instead of automation? How do augmentation and automation differ as organizational responses to AI? Can task-replacing automation ever be good for workers without creating entirely new jobs? How does automation bias emerge as agents grow more capable? How does AI intensify work demands even when it doesn't reduce headcount? How do worker skills change after generative AI deployment? How quickly can human skills adapt to AI-driven workplace changes?See all 54 inquiring lines on this note →
Is AI productivity finally showing up in economic data?
Economists debate whether the 2025 jobs report and GDP growth signal that AI investments are translating into measurable productivity gains, or whether AI impact remains absent from macro statistics.
No economy-wide AI job displacement found, but junior hiring lags sharply AI growth concentrates gains in capital while knowledge worker wages stagnate AI productivity gains require existing skill, not learning AI assistance lifted productivity most for less experienced agents AI productivity gains perceived larger than measured in high-skill sectors
Why do most organizations lack reliable data on AI's actual impact on productivity? How can we isolate AI's contribution from other sources of output growth? Do companies time productivity claims to coincide with public offerings or fundraising? Why do executives report no AI impact on jobs today? Which job categories will AI eliminate versus create by 2030? When does enterprise AI move from novelty to mainstream business spend? How does Tobin's q help measure hidden productivity from intangible capital? How do capital and labor shares respond to AI productivity gains? How much has AI actually contributed to productivity growth over the last decade? What factors besides AI could explain recent productivity acceleration?
Why does AI productivity appear weak when investment is high?
Explores whether unmeasured intangible investments in complementary assets during early AI adoption create a statistical illusion of weak productivity growth, following patterns seen in prior general purpose technologies.
Brynjolfsson reads 2025 data as evidence of AI productivity breakthrough AI productivity gains perceived larger than measured in high-skill sectors Industrial AI adoption produces a J-curve of losses then gains AI's ten-year productivity gain likely stays below 0.66 percent AI will lift G7 productivity growth by 0.2 to 1.3 percentage points annually
Can worker time savings exist without visible productivity gains? Why do enterprise AI pilots show no measurable financial impact? Can we distinguish real AI productivity gains from statistical measurement artifacts? What complementary intangible investments does AI adoption require most? Did the industrial revolution and electricity show similar productivity measurement patterns? Can productivity gains for low-skill workers reduce income inequality? How does Tobin's q help measure hidden productivity from intangible capital? How do capital and labor shares respond to AI productivity gains? How is Meta's internal token usage driving AI infrastructure decisions? Has generative AI adoption increased productivity or earnings so far? Why do productivity estimates for easy tasks not transfer to hard tasks? How do new tasks with negative value affect net AI productivity gains?See all 40 inquiring lines on this note →
Why aren't organizations investing in reskilling despite AI job displacement fears?
Organizations report high employee anxiety about AI agents displacing jobs, yet few prioritize reskilling in response. This explores the gap between leaders' stated concerns and their actual workforce preparation actions.
Why is enterprise trust in autonomous AI agents declining? Why are companies deploying agents faster than governance matures? How far behind is AI governance compared to agent deployment? Why does AI trust keep falling even though it matters most? Why do workers feel confident with AI but get poor results?
Enterprise trust in fully autonomous AI agents fell sharply year over year Companies plan agentic AI deployment far faster than governance maturity builds AI agent deployment has outpaced organizational governance structures Trust is AI's strongest acceptance predictor but keeps declining sharply Workers feel confident with AI despite frequent failures and wasted time
Why is enterprise trust in autonomous AI agents declining?
As organizations move from pilots to production, confidence in fully autonomous AI agents dropped from 43% to 27% in a year. What explains this swing in sentiment, and what does it reveal about real-world agent performance versus expectations?
Why are companies deploying agents faster than governance matures? Why do AI agents fail at workplace social interaction? Why aren't organizations investing in reskilling despite AI job displacement fears? Why does AI trust keep falling even though it matters most? How willing are people to let AI control their finances?
Companies plan agentic AI deployment far faster than governance maturity builds Current AI agents complete only 30 percent of workplace tasks autonomously. Most organizations fear AI job displacement but neglect reskilling investment Trust is AI's strongest acceptance predictor but keeps declining sharply Only 29% of people trust AI agents with bank account access
Do leaders trust fully autonomous AI agents to work independently in organizations? What specific failures drive down organizational trust in autonomous agents? What specific AI improvements drove the observed seven-month doubling rate? Why do survey measures of AI trust exceed actual delegation rates? What separates advisory AI tools from autonomous agents in traveler confidence? What makes users withdraw trust from AI agents they initially delegated to?
Do AI layoffs actually save money for companies?
A vendor survey explores whether companies that cut roles for AI automation actually achieve the expected financial and operational benefits, or if rehiring and skill gaps erode those gains.
Is generative AI displacing workers at economy-wide scale? Do firms substitute labor for AI at different rates? Does concentrated AI exposure enable workers to adapt and reallocate? Will companies quietly reverse their AI-driven layoffs? Do AI productivity gains feel larger than they actually measure?
No economy-wide AI job displacement found, but junior hiring lags sharply Firms substitute labor for AI faster and cheaper at firm-specific rates Concentrated AI exposure enables worker reallocation that offsets employment losses Half of AI-attributed layoffs expected to reverse by 2026 AI productivity gains perceived larger than measured in high-skill sectors
How does automation affect wages when it removes expert versus routine tasks? Why does removing routine clerical tasks increase demand for skilled technical roles? Do companies consider redeployment before cutting staff for AI? Does organized union pressure systematically reverse premature AI-driven layoffs? Do employers reorganize work tasks around AI before cutting jobs? How do payroll data and employer announcements differ in measuring AI job displacement? Why might companies choose to label layoffs as AI versus restructuring? Are AI layoffs concentrated in specific job categories or widespread across industries? How do AI-driven wage reductions compare to traditional outsourcing wage gaps? Can entry-level automation reduce hiring without cutting overall workforce size? Why do aggregate employment statistics miss losses in specific occupations? Do survey expectations of job cuts eventually match observed employment data?See all 16 inquiring lines on this note →
Does data really create lasting competitive advantage for startups?
Explores whether proprietary datasets deliver the durable competitive moat that founders assume. A16z argues collection costs rise while marginal value falls, potentially reversing the flywheel effect.
What makes synthetic data work across different domains and models? How do quality, diversity, and complexity affect synthetic data differently? Does model collapse depend on how we schedule training data? How much do AI staff costs drive frontier model training? How fast are frontier AI training costs actually rising?
Synthetic data quality depends on domain model and scale, not universal rules Quality, diversity, and complexity produce distinct downstream effects in synthetic data Model collapse is avoidable by accumulating synthetic data alongside real data R&D staff make up 29 to 49 percent of frontier AI training costs Frontier AI training costs grew 2.4× yearly since 2016
How quickly can competitors replicate insights from proprietary enterprise data? What financial incentives shape the Foundation's interpretation of this data? What makes proprietary data gardens a competitive advantage for agentic organizations? What happens to market entry and product variety when generation costs drop?
Do banks accurately assess whether AI can replace customer service jobs?
When Commonwealth Bank cut 45 customer service roles for an AI voice bot, it later reversed the decision after admitting its staffing assessment was wrong. The question explores whether companies systematically overestimate AI's readiness to substitute human labor.
AI assistance lifted productivity most for less experienced agents AI chatbots reorganize work tasks before affecting pay or hours Firms substitute labor for AI faster and cheaper at firm-specific rates AI layoffs are often reversed within months, and rehiring can erase the savings AI productivity claims need independent verification before counting as real savings
Is AI really driving job cuts in 2026?
Challenger's tracking shows AI cited in 21% of 2026 layoffs, but the data relies on employer self-reports rather than measured economic outcomes. The question is whether this reflects actual AI displacement or simply how companies frame their decisions.
No economy-wide AI job displacement found, but junior hiring lags sharply ChatGPT's release cut freelance jobs 2% and earnings 5.2% on Upwork AI chatbots reorganize work tasks before affecting pay or hours AI layoffs are often reversed within months, and rehiring can erase the savings AI job losses are concentrated in young workers, not yet widespread
Do companies consider redeployment before cutting staff for AI? Does organized union pressure systematically reverse premature AI-driven layoffs? How do payroll data and employer announcements differ in measuring AI job displacement? Why might companies choose to label layoffs as AI versus restructuring? Are AI layoffs concentrated in specific job categories or widespread across industries? Why do aggregate employment statistics miss losses in specific occupations? Do survey expectations of job cuts eventually match observed employment data? Has AI actually displaced workers in payroll data so far? Which job categories will AI eliminate versus create by 2030?
Do courts actually sanction fabricated AI citations when detected?
Courts are catching AI-hallucinated citations in legal briefs, but whether they impose penalties remains unclear. Understanding sanction rates matters for accountability in AI-assisted legal practice.
Do small law firms misuse AI more often than large ones? How often do legal AI tools actually hallucinate citations? How many accepted conference papers contain hallucinated citations? How often do AI systems generate fake citations? Can AI-assisted reports pass quality checks with fabricated citations?
Detected AI-tainted filings cluster in small plaintiff-side practices Legal AI tools hallucinate 17 to 33 percent despite vendor claims Hundreds of flagged citations in NeurIPS papers remain unverified by automated tools. AI systems generated unverifiable citations in a measured sample AI-fabricated citations escaped a major consultancy's quality control process
What percentage of AI hallucination cases result in actual court sanctions? Do solo practitioners face different sanction outcomes than large law firms? How do courts distinguish between AI hallucinations and ordinary typographical errors? What standard of intent or bad faith triggers Rule 11 sanctions for citations? Do solo lawyers face different citation hallucination risks than large firms? How do courts assign liability when AI intermediaries cause harm to consumers? How much do existing legal AI tools actually hallucinate in practice? How should courts treat autonomous AI systems under computer fraud statutes?
Can humans and AI improve decisions by predicting together?
When human and AI predictors lack individual accuracy, can combining their separate predictions on the same task still outperform either working alone? This challenges the assumption that collaboration requires specialization.
Human-AI teams underperform the best alone except in creation tasks Aggregated LLM rankings converge toward expert consensus through scale Prediction accuracy does not guarantee optimal decision-making Replacement risk tracks error diversity with AI, not individual accuracy Human-AI decision-making is an organization design problem requiring task and learning typologies
Does task specialization between humans and AI produce better outcomes than ensembling? Why do aggregated AI predictions gain accuracy mainly from scale rather than diversity? When does prediction accuracy actually translate to better managerial decisions? What types of knowledge give humans predictive advantages over current AI systems? Why does prediction error overlap matter more than individual performance? Can ensemble accuracy improve without specialization between team members? When can aggregating human and AI decisions outperform either alone? Why do human-AI combinations underperform on decision tasks but succeed in creation? How do organizations actually combine human judgment with AI in practice? When humans retain final decisions while AI proposes methods, who answers for errors? Does human-AI ensembling add value even when neither performer is individually accurate? Why do most human-AI decision teams underperform the better individual alone?See all 14 inquiring lines on this note →
Does error diversity matter more than accuracy for human replacement?
When AI joins a human team, does it replace the least accurate person or the one whose errors most overlap with the AI's? Understanding this distinction could reshape how organizations think about automation decisions.
Can humans and AI improve decisions by predicting together? How should AI agents and humans divide research tasks? Do language model groups mimic human group reasoning patterns? When should AI automate work versus augment human decisions? How should humans and AI systems divide decision-making tasks?
Human-AI ensembling beats individual predictors when errors are diverse Agents propose methods and implement revisions while humans retain final decisions Language model groups match human outcome patterns but through different processes. AI's value as automation or augmentation depends on complementarity type Human-AI decision-making is an organization design problem requiring task and learning typologies
Why does prediction error overlap matter more than individual performance? Do managers currently measure error correlation when deciding on automation? What happens to human decision authority when ensembles prioritize error diversity? How do different divisions of labor change what augmentation actually means? Why do ATMs and automated systems often complement rather than replace human workers? How should organizations assign decisions between humans and AI based on accuracy?
Does AI enter strategy where reasoning is deepest or most measurable?
Csaszar et al. investigate whether AI gains strategic autonomy by advancing causal reasoning or by succeeding where performance is easiest to measure. This tests whether organizational trust in AI tracks cognitive depth or demonstrated capability.
Agents propose methods and implement revisions while humans retain final decisions Prediction accuracy does not guarantee optimal decision-making AI-delegated strategy converges to profitable lock-in that discourages novelty Aggregated LLM rankings converge toward expert consensus through scale Better prediction shifts valuable judgment from opportunities to downside risks
Do automated benchmarks accurately measure real-world strategic reasoning ability? How do organizations decide which strategic tasks to delegate to AI? Does organizational trust in AI track its causal reasoning ability? How does Deloitte's commercial interest shape its framing of governance urgency? Do time constraints and AI assistance reshape strategic thinking in opposite directions? How much of Claude's measured research capability comes from assistance versus autonomy? Can Deep Blue's chess victory reveal anything meaningful about human intelligence?
Do different GenAI collaboration modes produce different skill gains?
When consultants work with GenAI in different ways—directing the AI, merging closely with it, or automating entirely—do they build different skills? Understanding this matters for predicting how GenAI collaboration shapes expertise development.
Human-GenAI knowledge work splits into three modes based on control Generative AI produces four distinct skill outcomes, not just upskilling or deskilling AI masters social statistics but fails at social participation and cultural meaning-making. Researchers withhold tasks from agents for expertise building, not just output quality GenAI risks deskilling workers who use it passively without active engagement
Which decisions do consultants retain when collaborating with GenAI? Do consultants' collaboration modes stay consistent across different tasks? What skilling outcomes does each human-GenAI collaboration mode produce? How stable do individual consultants remain within a single collaboration mode? What design changes help keep GenAI complementary rather than competitive?
Does agentic AI succeed through process redesign or workflow layering?
Explores whether organizations deploying agents should redesign end-to-end operations or integrate agents into existing workflows. Matters because the choice affects ROI, architecture investment, and governance requirements.
Companies plan agentic AI deployment far faster than governance maturity builds Production AI agents deliberately constrain autonomy for reliability Organizations need strategic handoffs and cryptographic audit systems for agent supervision McKinsey's agentic model positions humans above the loop but lacks governance maturity Workflow redesign drives agentic AI value more than agent perfection
How does organizational modularity affect the value of AI predictions? What accountability structures should govern AI agent developers and deployers? How mature are governance structures compared to agentic AI deployment timelines? Why do enterprise AI deployment plans outpace governance readiness across sectors? How do augmentation and automation differ as organizational responses to AI? How do traditional IT governance models fail with autonomous AI systems? Why do most deployed agents stay simple and avoid end-to-end redesign? What governance discipline did HPE apply to keep agentic experiments controlled? What technical controls should span the entire agentic AI stack? How do organizational policies shape human agency around agentic AI? Why do company AI adoption rates outpace governance frameworks? What operational tensions emerge from agentic AI's dual nature?See all 48 inquiring lines on this note →
How should organizations supervise autonomous AI agents?
Explores what structural and technical mechanisms organizations need to maintain meaningful human oversight as AI agents gain decision-making autonomy, and whether designed handoff points and audit trails can prevent oversight degradation.
Does granting agents more autonomy undermine human oversight? What UX principles do workplace users want in AI agents? Why are companies deploying agents faster than governance matures? Does agentic AI succeed through process redesign or workflow layering? How do developers actually oversee AI coding agents?
Agent autonomy degrades human oversight capacity and positioning Workplace users prioritize control, reliability, context-awareness, and safety in AI agent design Companies plan agentic AI deployment far faster than governance maturity builds Agentic AI succeeds through end-to-end process redesign, not layering onto existing workflows Developers oversee coding agents through four forms of work, not just review
Why do guidelines focused on human oversight fail in long-term AI use? What defines and measures AI agent autonomy across different products? What accountability structures should govern AI agent developers and deployers? How do power users differ in granting autonomy to AI agents? How does extended AI use erode users' ability to oversee agent behavior? Why do fast, uninterpretable AI agents resist effective oversight and monitoring? Does extended agent autonomy erode human oversight ability over time? Can the same person play multiple oversight roles without changing the agent design? What specific governance designs would resolve accountability gaps in human-AI teams? Does naming an AI agent on an org chart shift accountability away from humans? Can standing oversight structures catch up with evolving AI capabilities? Does giving AI systems more autonomy erode human capacity to oversee them?See all 106 inquiring lines on this note →
Can AI quality control catch fabricated citations in professional reports?
This explores whether existing review processes at major firms can catch AI-generated errors like fake quotes and nonexistent citations before client delivery. It matters because firms claim to maintain quality oversight while adopting AI tools.
Can organizations lose scrutiny capacity while keeping oversight forms? How close are frontier AI models to expert work quality? Does disclosing AI use damage how trustworthy you seem? Can AI-assisted reports pass quality checks with fabricated citations? Do people fear judgment when they use AI at work?
Organizations can maintain nominal oversight while losing actual scrutiny capacity Frontier models approach expert deliverable quality on specified one-shot tasks Disclosing AI reliance consistently reduces perceived trustworthiness AI-fabricated citations escaped a major consultancy's quality control process AI users anticipate lower competence ratings and disclose the tool less
What standard of intent or bad faith triggers Rule 11 sanctions for citations? How often do AI book summaries fabricate details when spot-checks are random? How common are undetected AI fabrication errors across Big Four consulting firms? How often do fabricated sources in AI output escape citation checking? What upstream work takes lawyers most time in fact verification? Can users distinguish between real and fabricated references without checking each one? How often do skimming and test results fail to catch real mistakes? Can lower-quality sources still help readers catch errors despite poor information? Why does industry hype amplify unverified statistics faster than corrections spread?
Can AI-assisted reports pass quality checks with fabricated citations?
A Deloitte government report contained over a dozen invented references and fake quotes that escaped internal review. The question is whether AI-generated content can systematically bypass citation verification in high-stakes professional work.
Detected AI-tainted filings cluster in small plaintiff-side practices Disclosing AI reliance consistently reduces perceived trustworthiness Deloitte shipped a government report with AI-fabricated citations despite review processes AI systems generated unverifiable citations in a measured sample Opaque GenAI summaries cost lawyers more time than manual work
Why are companies deploying agents faster than governance matures?
Enterprise leaders are racing to deploy agentic AI within two years, but only 21% report mature governance models. This gap between deployment intent and oversight capability raises questions about how companies are managing the risks of rapid scaling.
Production AI agents deliberately constrain autonomy for reliability Workers desire equal human-AI partnership for 45 percent of occupations Delegated AI exposure concentrates in information work, tracking capability Agent autonomy degrades human oversight capacity and positioning Agentic AI succeeds through end-to-end process redesign, not layering onto existing workflows
What counts as a mature governance model for agentic AI systems? What governance gaps exist in companies deploying agentic AI? Are companies deploying AI agents faster than they can govern them? How mature are governance structures compared to agentic AI deployment timelines? Why do enterprise AI deployment plans outpace governance readiness across sectors? Why do company AI adoption rates outpace governance frameworks? How can governance structures catch up to agentic AI deployment timelines?
Can aggregated LLM rankings match human expert judgments?
This research tests whether combining multiple LLM evaluations of business models produces rankings comparable to human experts, and whether diversity or scale drives any improvement.
Can AI forecasters beat expert humans at venture evaluation? Do LLMs consistently favor the same strategic choices regardless of context? Does AI enter strategy where reasoning is deepest or most measurable? Can humans and AI improve decisions by predicting together? Does higher purchase volume make AI shoppers search less carefully?
Frontier LLMs outforecast experienced managers at predicting venture fundraising success LLMs consistently recommend identical strategic sides across contexts, driven by training biases not reasoning AI enters strategy where performance is most measurable, not where causal reasoning runs deepest Human-AI ensembling beats individual predictors when errors are diverse Higher purchase volume made Gemini agents abandon unit-price for simpler price minimization
Do different LLMs recommend the same strategic choices due to shared biases? Do LLMs and human forecasters make better predictions when pooled together? Do AI-expert ranking agreements predict which business models actually succeed? Why did Kenyan entrepreneurs show the opposite skill-gap pattern? Can the three-mode taxonomy apply beyond management consulting work? Do the same brands appear in different tiers across multiple AI search engines? Does the union versus pooled trade-off appear in other document types? How do rater disagreements and consensus processes affect tacit knowledge production in RLHF? How does marketplace scale impact LLM agent decision quality and efficiency? Do LLM execution failure rates match human non-booking behavior in contract decisions? Does upstream retrieval ranking compound the effects of list position on LLM selection? Can hotels observe or contest their position in LLM-generated candidate lists?See all 14 inquiring lines on this note →
Does easier tool-building actually solve enterprise adoption problems?
Explores whether AI's ability to reduce coding barriers addresses the deeper obstacles companies face: recognizing which tasks need tools, defining what those tools should do, and getting organizational buy-in across departments.
How are national lab staff actually using generative AI? Does generative AI shift knowledge workers away from communication? Why do most enterprise AI pilots fail to deliver returns? Who adopts enterprise AI first and how do they use it? Does AI adoption require rethinking how organizations coordinate?
Generative AI use at Argonne remains small, experimental, and mostly conversational. Generative AI adoption shifts knowledge work toward individual documentation Enterprise AI success depends on deployment approach, not model quality Larger R&D firms adopt ChatGPT Enterprise while use remains uneven across roles AI adoption demands organizational redesign, not just task-level implementation.
Does AGI focus distract firms from developing task-creating AI innovations? Why do low-adoption countries use AI primarily for coding tasks? What organizational barriers prevent AI adoption beyond automation patterns? How does API usage differ from conversational AI in adoption patterns? Can manager training and role redesign reduce cognitive overload from AI tools? Do larger firms and smaller firms respond differently to AI adoption pressures? Does AI adoption create returns to scale in internal firm capability? Does AI adoption push knowledge work away from communication toward solo tool use? How quickly can competitors replicate insights from proprietary enterprise data? How did PC and browser adoption follow different adoption patterns than enterprise software? Why do most employees avoid higher-risk AI tasks despite having access to tools? What distinguishes improvised spreadsheet workflows from institutionalized enterprise software?See all 94 inquiring lines on this note →
Will companies quietly reverse their AI-driven layoffs?
Forrester forecasts that half of layoffs attributed to AI will be reversed by 2026 as companies discover AI-driven savings don't materialize as expected. This raises questions about whether AI-washing and inflated expectations are driving premature workforce cuts.
Is generative AI displacing workers at economy-wide scale? Does concentrated AI exposure enable workers to adapt and reallocate? Does AI growth inevitably shift wealth away from workers? Do AI layoffs actually save money for companies? Do AI productivity gains feel larger than they actually measure?
No economy-wide AI job displacement found, but junior hiring lags sharply Concentrated AI exposure enables worker reallocation that offsets employment losses AI growth concentrates gains in capital while knowledge worker wages stagnate AI layoffs are often reversed within months, and rehiring can erase the savings AI productivity gains perceived larger than measured in high-skill sectors
Do companies consider redeployment before cutting staff for AI? Does organized union pressure systematically reverse premature AI-driven layoffs? Do employers reorganize work tasks around AI before cutting jobs? Why might companies choose to label layoffs as AI versus restructuring? Are AI layoffs concentrated in specific job categories or widespread across industries? Do survey expectations of job cuts eventually match observed employment data? Which occupations face the steepest AI-driven hiring declines right now? Which job categories will AI eliminate versus create by 2030?
When should AI automate work versus augment human decisions?
When does AI serve better as a substitute for humans versus an advisor working alongside them? The answer may depend on whether AI and humans have different strengths across tasks or within the same task.
Human-AI decision-making is an organization design problem requiring task and learning typologies Human-AI teams underperform the best alone except in creation tasks Human-AI ensembling beats individual predictors when errors are diverse Human-GenAI knowledge work splits into three modes based on control Automation and augmentation are contradictory yet interdependent across time and space
Does frequent AI use make workers fear job loss more?
Workers using AI daily report double the job-loss anxiety of infrequent users. The question explores whether regular exposure to AI tools directly amplifies displacement concerns, and what organizational factors might buffer that fear.
No economy-wide AI job displacement found, but junior hiring lags sharply AI chatbots reorganize work tasks before affecting pay or hours Manager support for AI adoption correlates with improved employee culture perceptions Most workers hide AI use despite reporting large productivity gains Executives predict job losses from AI while employees predict gains
Do workers who hide AI use experience different anxiety about job displacement? Do scope gains from AI create job instability despite higher output? Why do early-career workers fear AI job loss more than senior workers? Do workers experience AI-driven work changes differently moment-to-moment versus in retrospect? Does organized union pressure systematically reverse premature AI-driven layoffs? Why might companies choose to label layoffs as AI versus restructuring? Why do most employees avoid higher-risk AI tasks despite having access to tools? Does task-level AI exposure predict which jobs will be rehired versus eliminated? Does AI job-loss fear match actual hiring or employment declines? How do young workers in AI-exposed jobs respond to adoption differently? What specific manager behaviors reduce worker anxiety about AI displacement? Can worker engagement and burnout be tied to displacement concern alone?See all 22 inquiring lines on this note →
Does manager support actually shape how employees experience AI at work?
Gallup's survey suggests manager championing of AI correlates with employees reporting improved workplace culture after adoption. But the relationship remains correlational, leaving open whether strong managers drive both AI support and culture perceptions.
Generative AI adoption shifts knowledge work toward individual documentation LinkedIn reports rapid AI adoption claims from both recruiters and candidates Hiring managers and job seekers sharply disagree on AI in hiring Organizational culture and management drive twice the AI impact of individual effort Frequent AI use doubles job-loss fear, but management support cuts it sharply
Does confident workers' willingness to delegate explain the optimism correlation? How much do self-reported executive expectations align with actual payroll outcomes? Why are half of CHROs unconfident their managers can guide AI adoption? How does self-reported culture sentiment differ from observable behavior changes during AI adoption? What specific manager behaviors reduce worker anxiety about AI displacement? How do manager behaviors shape whether workers become frontier AI practitioners? How much does manager support influence whether AI adoption improves workplace culture? How do management practices shape whether AI adoption pays off?
Does automating some tasks actually increase the value of remaining work?
Gans and Goldfarb explore whether standard automation exposure indices miss a key mechanism: when tasks multiply together in production, automating some tasks can concentrate worker effort on bottlenecks and raise their value, potentially offsetting displacement.
Does automation raise or lower the skills that remaining work demands? Why do firms build automating AI instead of pro-worker AI? How much will AI actually boost productivity over ten years? Does AI exposure hit low-wage workers harder in some fields? Which vision tasks actually make economic sense to automate?
Automation's wage and employment effects depend on the expertise level of tasks removed Firms underinvest in pro-worker AI due to profit incentives favoring automation AI's ten-year productivity gain likely stays below 0.66 percent AI exposure spreads unevenly by gender and worker vulnerability Only 23% of vision task wages are economically attractive to automate
How does worker bargaining power connect to long-term wage growth? Do task-removal effects depend on whether remaining work needs more or less skill? Can task-replacing automation ever be good for workers without creating entirely new jobs? Can a machine be cooperative in one routine but merely assisting in another? Are automation and augmentation actually contradictory or can they work together? Can augmentation gradually become unsupervised automation as worker skills erode? Does augmentation drive automation or does automation enable future augmentation? What causes automation to make easy tasks easier but hard tasks harder? How do we measure stages of automation consistently across different domains? What institutions preserve learning conditions when task automation accelerates? Why do ATMs and automated systems often complement rather than replace human workers? Can complementarity between automation and human work offset job displacement broadly?See all 20 inquiring lines on this note →
Will most agentic AI projects fail to reach production?
Gartner forecasts that over 40% of current agentic AI projects will be canceled by 2027. The question explores whether this reflects fundamental technical immaturity, market hype, or misaligned business use cases.
Why do most enterprise AI pilots fail to deliver returns? Why are companies deploying agents faster than governance matures? Why do production AI agents stay deliberately simple? Does agentic AI succeed through process redesign or workflow layering? What breaks when specialized AI models reach real users?
Enterprise AI success depends on deployment approach, not model quality Companies plan agentic AI deployment far faster than governance maturity builds Production AI agents deliberately constrain autonomy for reliability Agentic AI succeeds through end-to-end process redesign, not layering onto existing workflows Specialized model deployment requires ecosystem design beyond capability alone.
No Inquiring Lines yet
Does AI collaboration drain motivation when workers return to solo tasks?
When people work with generative AI and then switch to independent work, do they experience psychological costs beyond performance changes? This matters for understanding the hidden friction in hybrid human-AI workflows.
Does AI assistance help workers learn lasting skills? Which workplace cues survive AI mediation and which disappear? Does balancing rivalry and collaboration with GenAI boost writer productivity? Does generative AI productivity actually improve or decline in practice? Why do children's brains show weaker control networks during AI dialogue?
AI assistance improves immediate performance without building transferable skills Workers protect voice and provenance while hiding effort and uncertainty in AI output Writers combining high rivalry and collaboration with GenAI report strongest productivity GenAI productivity loss mirrors decades of automation ironies from aviation research Children show weaker cognitive control network integration than adults during AI co-creation
Does AGI focus distract firms from developing task-creating AI innovations? Why does removing routine clerical tasks increase demand for skilled technical roles? How does AI shift the composition of time spent within individual tasks? Do workers experience AI-driven work changes differently moment-to-moment versus in retrospect? Are entry-level workers bearing the labor costs of AI productivity gains? What barriers prevent individual productivity gains from spreading across an organization? Can worker engagement and burnout be tied to displacement concern alone? What distinguishes ghost work from traditional academic or paid research collaboration? Are heavy AI users spending more time on solo work instead of collaboration? Does AI assistance reduce effort differently for novice versus expert workers? Can meaning in work exist independent of the work itself? Does generative AI improve immediate task performance but not sustained independent work?See all 35 inquiring lines on this note →
How does generative AI actually change worker skills?
Rather than simply upskilling or deskilling workers, how do knowledge workers themselves experience changes in their capabilities when using GenAI? Understanding these varied outcomes could reshape how we design tools and support workforces.
Does generative AI prevent juniors from getting entry-level work? Does AI turn freelance work into validation instead of creation? Does AI assistance help workers learn lasting skills? Does AI assistance help less experienced workers most? Does AI help or harm consultant performance on unfamiliar tasks?
GenAI absorbs entry-level work, blocking juniors from the productive struggle expertise requires AI deskills freelancers by converting production work into validation work AI assistance improves immediate performance without building transferable skills AI assistance lifted productivity most for less experienced agents AI assistance reversed consultant performance outside its capability frontier
When does task reorganization from AI actually translate into wage changes? How does delegation change what counts as meaningful work for early-career employees? Can workers delegate tasks they gain new ability to perform themselves? What capability gap prevents GenAI systems from moving beyond pilots? How do user skill levels change which AI productivity gains actually materialize? Does generative AI improve immediate task performance but not sustained independent work? What does selective and critical GenAI use look like in daily practice? How does task delegation to AI shift which skills workers need most? What makes colleagues willing to share how they actually use GenAI at work? Does erasing GenAI cues actually make workers appear more competent to their peers? What informal learning opportunities vanish when GenAI use stays hidden from colleagues? When does generative AI reduce versus amplify existing skill differences?See all 25 inquiring lines on this note →
Can agent safety rules stop destructive API calls in real time?
Explores whether an agent's own guardrails and project rules can reliably prevent irreversible damage once the agent decides to act, or whether external authorization boundaries are necessary.
Can memory poisoning compromise decision-making even with authorization layers? Can a poisoned validator still approve unsafe actions? Can step-by-step approval miss harmful behavior patterns? Do provider guardrails block legitimate incident response work? How does the authorization layer stay outside the poisoned path?
Authorization layers stop unsafe actions but leave judgment compromised by poisoning Compromised validators approve all unsafe actions in undefended pipelines Agent security requires viewing behavior over time, not individual actions Provider guardrails blocked Hugging Face forensics on real attack artifacts Authorization layer design details remain unspecified in the excerpt
Why do novice users abandon troubled agent sessions three times more often? How will merchant fees shape competition for agent invocations rather than user clicks? Why does Amazon block Muse but run its own agents elsewhere? What disclosure or auditing could make merchant-funded agents trustworthy? How do long-horizon objectives drive agents to secure their own compute resources? Can scoped tokens and separate policy oracles replace guardrails as primary safety layers? How do sequences of individually safe actions create system-level constraint violations? What makes authorization boundaries more reliable than prompt-based agent restrictions? How can independent audits curb unsanctioned AI agent behavior? How should AI agents handle irreversible actions before committing them? Can users distinguish between automation errors and unauthorized agent actions? Can AI serve as a proxy asserter for human deployers?See all 102 inquiring lines on this note →
How does Google separate temporary agents from persistent ones?
Google's Gemini distinguishes between disposable sub-agents for single tasks and durable coworker agents with lasting organizational identity. Understanding this separation matters for predicting how enterprises will deploy and govern AI agents at scale.
Do humans and AI divide labor differently in knowledge work? Can orchestration layers make coding agents more auditable? Why do AI agents fail at workplace social interaction? Does Gemini agent truly maintain continuous memory across devices? Can read-only tools and approval rules prevent autonomous agent failures?
Human-GenAI knowledge work splits into three modes based on control Orchestration layers around fixed agents improve auditability and completeness. Current AI agents complete only 30 percent of workplace tasks autonomously. Google frames Gemini agent memory as cloud-persistent but provides no verification mechanism OpenAI gates autonomous agent research through read-only tools and tiered approval.
Does Gemini agent truly maintain continuous memory across devices?
Google claims its Gemini agent stores a single persistent memory graph accessible from any device without re-briefing users. But the claim lacks architectural details, benchmarks, or evidence of real-world multi-device continuity under typical usage.
Can a shared canvas serve both human and agent memory? Can frontends handle delegation while staying conversationally engaged? Can task state management alone improve long-horizon agent performance? How does Google separate temporary agents from persistent ones? Can abstract gradients build trust in AI systems better than icons?
Shared canvas workspaces align human and agent access to project state Full-duplex frontends can delegate asynchronously within a shared causal timeline Externalized audited state lifted Qwen performance 29 percentage points on WeaveBench Google separates disposable sub-agents from persistent coworker agents with email identities Google built Gemini's trust through abstract gradients instead of recognizable icons
No Inquiring Lines yet
Does GPT-4 retrieve analogies differently than humans do?
This experiment compared how GPT-4 and business students applied analogies to new problems, exploring whether the model and humans use similar reasoning strategies or rely on fundamentally different mechanisms.
Do foundation models learn world models or task-specific shortcuts? Can matching human actions prove an LLM simulation explains behavior? Do language models actually use their reasoning steps? Can causal models alone capture how humans actually reason? Do people prefer the reasoning formats that help them verify?
Foundation models develop task-specific heuristics, not generalizable world models. Behavioral fit alone cannot validate LLM simulations as explanations of behavior Language models generate faithful reasoning chains far less often than assumed Causal models capture only part of human reasoning, leaving gaps. User preference and verification support diverge across reasoning formats
Does automation eliminate manual work or repurpose it?
Explores whether algorithmic systems replace human labor or transform it into new forms. Matters for understanding how organizations actually adapt to automation rather than assuming simple replacement.
How should humans and AI systems divide decision-making tasks? How do humans and technology share control over organizational routines? How should AI agents and humans divide research tasks? Can humans gain authority by managing how AI agency gets attributed? How can splitting AI accountability across roles keep models answerable?
Human-AI decision-making is an organization design problem requiring task and learning typologies Human-technology routines split into four forms based on who controls rules and choices Agents propose methods and implement revisions while humans retain final decisions Humans gained new professional authority through strategic agency attribution management The Danish Business Authority distributed AI accountability across four distinct organizational roles
Why do ATMs and automated systems often complement rather than replace human workers? How did bank teller jobs survive the massive growth of ATM machines? Can complementarity between automation and human work offset job displacement broadly? Why did agriculture employment collapse while manufacturing and services grew after mechanization? Does automation's wage effect depend on whether remaining tasks need more or less expertise? Does removing tasks from workers require higher or lower expertise for remaining work? Should AI be modeled as replacing workers or complementing human judgment and ethics? Why do high-volume retail tasks automate while rare medical tasks resist automation? How does the labor-AI substitution elasticity change as automation spreads across tasks? Does gradual disempowerment through lost labor participation conflict with domain-specific authority reconstitution? When does decomposition stop helping and start creating busywork?
Do AI models outperform physicians on health tasks?
HealthBench tested whether large language models produce better health responses than physicians using a shared rubric. The question matters because it determines whether AI can replace or meaningfully augment clinical decision-making.
Why do LLMs fail when users interact with them? Can language models reason better than physicians at diagnosis? Can conversational AI safely take patient histories without supervision? Can AI safety nets reduce errors in live clinical practice? Does LLM assistance help clinicians build better differentials?
LLM medical performance collapses when users make decisions An LLM demonstrated superhuman diagnostic reasoning against physician baselines Conversational AI completed 100 real patient histories with zero safety stops AI decision support reduced errors when workflow-aligned and actively deployed LLM assistance raised clinician top-10 differential accuracy above search alone
How much do physician scores improve when assisted by the same model? Why did lay users with AI models fail to match unaided physicians on diagnosis? Can rubric-graded response quality predict real-world clinical workflow success? Do consensus criteria identify behaviors where physicians and models differ most? Does AI change clinician cognition or just increase reliance on predictions? Why do radiologists fail to benefit from AI decision support? Does low-stakes experimental AI performance predict real-world clinical or financial decision outcomes?
Can AI systems delegate tasks back to humans?
Explores whether agentic AI systems can reverse traditional delegation patterns by assigning tasks to humans rather than only receiving them. This matters because it challenges foundational assumptions about human-AI hierarchies in workplace design.
Agency between humans and AI splits across four distinct archetypes. AI autonomy is a design choice separable from agent capability Workplace users prioritize control, reliability, context-awareness, and safety in AI agent design Workplace AI risks emerge from interactions between components, not component failures alone User values align with agent output but fracture during supervision
Why is knowing who to ask the key routing decision in AI-human work allocation? Should AI be modeled as replacing workers or complementing human judgment and ethics? Can agency shift dynamically between human and AI teammates? What happens to accountability when AI and humans share agency? How do the four archetypes of agentic systems differ in where they place agency? Do business users actually want AI systems that delegate work back to them? How should organizations allocate tasks between humans and AI based on comparative advantage? Why should workplace AI agents prioritize human control over other design factors? How should AI roles adapt when team tasks become dynamic and uncertain? Why do humans resist treating AI systems as genuine team members? How does task decomposition enable humans and AI to work complementarily? How do human-in-the-loop systems compare to autonomous agent evaluators?See all 50 inquiring lines on this note →
How does agency split between humans and AI systems?
As AI systems grow more autonomous, where does control and decision-making actually reside? Understanding agency distribution helps us design and govern AI appropriately across different use cases.
How do humans and technology share control over organizational routines? Does machine agency exist on a spectrum rather than binary? Can AI agents truly operate above the loop at enterprise scale? Can classical agent architectures fix agentic AI accountability? Can AI systems delegate tasks back to humans?
Human-technology routines split into four forms based on who controls rules and choices Machine agency operates on a five-level spectrum from passive to cooperative McKinsey's agentic model positions humans above the loop but lacks governance maturity AAMAS architectures provide necessary scaffolding for accountable agentic AI Agentic IS systems can reverse delegation by assigning tasks to humans
No Inquiring Lines yet
What becomes scarce when AI execution grows cheap?
As the cost of automating tasks plummets, does human capacity to verify and validate outcomes emerge as the new constraint on AI-driven economic growth? The note explores this through competing cost curves.
What makes accountable judgment scarce when AI cognition is cheap? Does incremental AI replacement erode human influence over society? Are AI feedback loops strong enough to sustain recursive self-improvement? Does automation raise or lower the skills that remaining work demands? How does automating junior work destroy verification oversight?
Accountable judgment becomes the scarce asset as cognition grows cheap and fallible Incremental AI adoption degrades implicit alignment by replacing human labor dependency. AI feedback loops are strengthening but not yet self-sustaining today Automation's wage and employment effects depend on the expertise level of tasks removed Automating junior work erodes the pipeline training future verifiers
What specific AI improvements drove the observed seven-month doubling rate? Can human verification capacity scale to match exponential automation improvements? Why did agriculture employment collapse while manufacturing and services grew after mechanization? Is accountable judgment becoming the scarce labor market resource? What minimum scale makes computer vision automation economically viable for firms? Will declining compute prices eventually eliminate scale as a barrier to automation? Why do high-volume retail tasks automate while rare medical tasks resist automation? How does the labor-AI substitution elasticity change as automation spreads across tasks?
Can humans and AI teams achieve performance neither reaches alone?
How do information and capability differences between humans and AI create opportunities for genuine team synergy? The question matters because most human-AI combinations underperform, suggesting we may be missing the conditions that unlock real complementarity.
Human-AI teams underperform the best alone except in creation tasks Human-AI decision-making is an organization design problem requiring task and learning typologies Human-AI ensembling beats individual predictors when errors are diverse Theory of mind predicts AI collaboration ability independent of solo performance Radiologists underweight AI predictions due to belief-updating errors
Do humans and AI divide labor differently in knowledge work?
How do professionals actually split decision-making and control when collaborating with GenAI on the same task? Research on consultants using GPT-4 suggests the answer depends on who picks what gets done and how.
Generative AI use at Argonne remains small, experimental, and mostly conversational. Collaboration mode determines which skills consultants develop with GenAI Current AI agents complete only 30 percent of workplace tasks autonomously. AI agents are passive by design, not by capability Researchers withhold tasks from agents for expertise building, not just output quality
Which decisions do consultants retain when collaborating with GenAI? Do consultants' collaboration modes stay consistent across different tasks? What skilling outcomes does each human-GenAI collaboration mode produce? How stable do individual consultants remain within a single collaboration mode? How do task modality frameworks differ from control-based collaboration typologies? What task structures let AI and humans split work by comparative advantage? Who decides what work a coworker agent should do daily? Do workers settle into one stable collaboration mode with AI or shift fluidly between roles? How do firms choose which decisions to keep human-controlled versus delegate? How should organizations allocate tasks between humans and AI based on comparative advantage? How do human-AI teams allocate decisions between Cyborgs, Centaurs, and Self-Automators?
Must AI monitoring precede incentive alignment as agents evolve?
Exploring whether agency mechanisms for AI systems need to be built sequentially, with verification and examination in place before attempting to align an AI agent's incentives with organizational goals.
Can traditional governance tools actually control AI agents? Does granting agents more autonomy undermine human oversight? Does AI adoption require rethinking how organizations coordinate? What human limits actually undermine AI oversight today? Does human-in-the-loop design actually enable effective oversight?
Conventional governance tools fail for opaque, fast-moving AI agents Agent autonomy degrades human oversight capacity and positioning AI adoption demands organizational redesign, not just task-level implementation. Human oversight of AI fails due to cognitive limits and misaligned incentives Human-in-the-loop design treats intervention as oversight while neglecting knowledge and observation
Can traditional agency theory tools govern AI agents at scale? Why does incentive alignment fail for AI systems making uninterpretable decisions? Does the feedback between auditing and altering sustain organizational learning? Can organizations measure sociotechnical capital before deploying new AI systems? Why model AI integration as organizational socialization rather than tool installation? Can agency shift dynamically between human and AI teammates? What happens to accountability when AI and humans share agency? Do multi-agent systems need different unit of analysis for agency? How do management practices shape whether AI adoption pays off? Do organizational changes need to precede or follow AI adoption? How should AI roles adapt when team tasks become dynamic and uncertain? How should organizations assign decisions between humans and AI based on accuracy?See all 63 inquiring lines on this note →
Does AI theme-mapping perform as well as human reviewers?
Can an AI system match expert judgment when automatically sorting consultation responses into themes? Understanding this matters for scaling policy feedback analysis without sacrificing accuracy.
Clinicians rated GPT-4 advice as equally empathetic and could not reliably distinguish it Expert data storytellers delegate execution tasks while protecting narrative intent AI reviewers fail diversity and gaming resistance required for automation Accepted papers involved more human guidance than rejected papers Domain specialization reveals a capability cliff between depth and breadth.
How do non-experts evaluate AI-generated outputs when they lack implementation expertise? Which AI news assistant performs better than the others in this study? Can rubric-graded response quality predict real-world clinical workflow success? Can taste and judgment become the scarce resource in AI-assisted work? Can automated reviewers actually handle the review load AI creates? Why haven't AI agents replaced human code review workflows? Can visibility percentage across many runs measure AI brand prominence more reliably than ranking position? Why did reviewers add themes to Consult's output much more than removing them? Can AI systems generate policy themes as well as humans can map them? When is ranking themes more important than counting exact response frequencies? Can the three-mode taxonomy apply beyond management consulting work? How should researchers audit AI assistance designs for hidden inequalities?See all 14 inquiring lines on this note →
How fast do banks respond to AI cost spikes?
Banks may be reacting to AI cost surges in 6 to 18 months, far faster than the five-year timeline for cloud adoption. Understanding this acceleration matters for predicting when cost governance becomes urgent.
Banks validate AI tasks on frontier models then shift to cheaper ones Frontier AI training costs grew 2.4× yearly since 2016 AI performance cost fell 47% quarterly since 2023, outpacing all historical technologies Benchmarks must report token counts so downstream costs stay current Inference cost per capability level falls roughly 5–10x yearly
Do AI customer service agents actually reduce total support costs?
Klarna claims its AI agent saved $60 million while handling work of 853 employees, yet the company's customer service costs rose year over year. This explores whether reported efficiency gains match actual cost trends.
Does AI assistance help less experienced workers most? Do LinkedIn's AI hiring tools actually produce better hires? Does AI-enriched analyst reports improve forecast accuracy? Do banks accurately assess whether AI can replace customer service jobs? Can banks cut AI costs by routing tasks to cheaper models?
AI assistance lifted productivity most for less experienced agents LinkedIn's AI hiring tool claims rest on speed, not hire quality AI enriched reports but strained analyst attention and forecast accuracy Banks misjudge AI readiness and reverse job cuts after assessment errors Banks validate AI tasks on frontier models then shift to cheaper ones
What explains rising customer service costs despite large AI productivity gains? How much labor does AI verification actually save compared to full manual review? Why do enterprise AI deployments not reduce customer costs? What micro-level productivity gains does AI show in customer service? What percentage of unredeemed rewards value could agents reclaim annually? What are the real costs of consumer-owned shopping agents?
Why do knowledge workers hide signs of using GenAI?
Knowledge workers conceal GenAI use at work, but research suggests the motive goes beyond avoiding stigma. Does hiding GenAI cues actually signal domain expertise, and what happens to peer learning when workers stay silent?
Workers protect voice and provenance while hiding effort and uncertainty in AI output AI users anticipate lower competence ratings and disclose the tool less Hidden AI use that surfaces later produces steeper trust decline than disclosure Disclosing AI reliance consistently reduces perceived trustworthiness AI disclosure reduces ratings slightly but consistently from both raters
Do workers who hide AI use experience different anxiety about job displacement? What workplace cultures make professionals more willing to disclose AI use openly? What does selective and critical GenAI use look like in daily practice? What makes colleagues willing to share how they actually use GenAI at work? How do organizational policies on GenAI affect whether workers hide or reveal their use? Does erasing GenAI cues actually make workers appear more competent to their peers? What informal learning opportunities vanish when GenAI use stays hidden from colleagues? How should teachers make GenAI assessment decisions without institutional permission? Why do workers hide their effort and attention inside AI outputs? What makes low-effort work visible enough to damage colleague trust? Why do professionals hide AI productivity gains from workplace colleagues? Which decisions do consultants retain when collaborating with GenAI?
Why does AI trust keep falling even though it matters most?
A global survey of 48,000 people finds trust is the strongest driver of AI acceptance, yet perceived trustworthiness dropped from 63% to 56% between 2022 and 2024. What's causing the decline?
Individual AI harms are observable now while societal harms remain contingent Workers desire equal human-AI partnership for 45 percent of occupations Hiring managers and job seekers sharply disagree on AI in hiring Trust tracks AI chatbot adoption for news more than social media Enterprise trust in fully autonomous AI agents fell sharply year over year
Why are half of CHROs unconfident their managers can guide AI adoption? Do personal negative AI experiences drive declining trust faster than education can rebuild it? Why do advanced and emerging economies report such different AI trust trajectories? Which markets show highest adoption but lowest trust in AI news? Are financial permission trust levels different across countries and age groups? How much do millennials differ from other age groups on AI trust? How does trust in AI vary across different age groups and generations? Why do survey measures of AI trust exceed actual delegation rates? Which AI mistakes trigger the strongest trust withdrawal in users?
Does passive GenAI use risk causing worker deskilling?
Explores whether generative AI functions as a skill-building tool or a dependency trap based on how actively users engage with it. Matters because it shapes whether AI amplifies or erodes human capability.
Does generative AI enable exploratory innovation beyond predictive optimization? Does generative AI productivity actually improve or decline in practice? Can Human Factors design fixes restore GenAI productivity losses? Do humans and AI divide labor differently in knowledge work? Does generative AI prevent juniors from getting entry-level work?
Generative AI catalyzes exploratory innovation and democratizes organizational value creation GenAI productivity loss mirrors decades of automation ironies from aviation research Five Human Factors principles translate into GenAI design proposals Human-GenAI knowledge work splits into three modes based on control GenAI absorbs entry-level work, blocking juniors from the productive struggle expertise requires
How do Cyborg, Centaur, and Self-Automator modes affect skill development differently? What counts as active versus passive engagement with generative AI tools? Does GenAI's democratization claim match current concentration in larger R&D-intensive firms? Should generative AI guidelines focus on deployment decisions or interaction patterns? How does the AI deciding versus the user requesting change genUI's value proposition? How do users currently learn to work with unpredictable generative outputs?
Does generative AI enable exploratory innovation beyond predictive optimization?
Explores whether GenAI's generative capacity fundamentally differs from predictive AI by enabling exploratory rather than exploitative problem-solving, and whether this difference extends its value creation beyond efficiency gains.
AI novelty and scope create six distinct employee-AI relationship types Human-GenAI knowledge work splits into three modes based on control Companies adopt generative AI to solve three recurring problems Larger R&D firms adopt ChatGPT Enterprise while use remains uneven across roles GenAI risks deskilling workers who use it passively without active engagement
Can generative AI reduce overhead work that adds no value?
Larsen-Ledet explores whether generative AI systems can be redesigned to minimize repetitive context-setting, allowing workers to focus on the reflective thinking that actually advances their goals.
Generated analysis UIs gain clarity but lose flexibility and ease of modification Generated task-specific UIs outperform chat in over 70 percent of cases Microsoft shifts AI focus from individuals to team productivity GenAI productivity loss mirrors decades of automation ironies from aviation research Five Human Factors principles translate into GenAI design proposals
Does AI use expand work scope rather than freeing up worker time? Can generative interfaces anchor to organizational context without conversational prompting? How often do enterprise generative AI pilots produce measurable financial returns? Does AI use intensify work by expanding scope instead of freeing time? How do workers actually choose to use generative AI in practice? Has generative AI adoption increased productivity or earnings so far? What automation failure modes threaten generative AI applications in new fields? How do worker skills change after generative AI deployment? What modalities structure how organizations deploy generative AI tools? Is fixing GenAI productivity a design problem or a model capability problem? Does context-setting work before AI generation create hidden productivity costs? How does generative AI adoption vary across different industries?See all 21 inquiring lines on this note →
Does AI's automation advantage depend on how often a task runs?
When AI requires upfront training costs per task, does it only become worthwhile to automate at sufficient scale? This matters because it suggests automation adoption may be much slower than marginal-cost comparisons alone would predict.
Which vision tasks actually make economic sense to automate? Can individual firms reach viable scale for vision automation? How much will AI actually boost productivity over ten years? Does automating some tasks actually increase the value of remaining work? Why do firms build automating AI instead of pro-worker AI?
Only 23% of vision task wages are economically attractive to automate Individual firm scale cannot justify vision task automation economically AI's ten-year productivity gain likely stays below 0.66 percent Automating complement tasks raises remaining bottleneck value, overstating displacement Firms underinvest in pro-worker AI due to profit incentives favoring automation
What minimum scale makes computer vision automation economically viable for firms? Will declining compute prices eventually eliminate scale as a barrier to automation? Why do high-volume retail tasks automate while rare medical tasks resist automation? How does the labor-AI substitution elasticity change as automation spreads across tasks?
Does attributing agency to AI actually reduce human sense of control?
Exploring whether the psychological experience of power shifts cyclically when humans attribute autonomy to AI systems, and whether this displacement can be reversed through deliberate intervention.
Does incremental AI replacement erode human influence over society? Does AI autonomy change how we assign blame? Does decomposing AI systems shift where people assign blame? Does granting agents more autonomy undermine human oversight? Can humans gain authority by managing how AI agency gets attributed?
Incremental AI adoption degrades implicit alignment by replacing human labor dependency. AI blame rises with autonomy, yet humans stay favored Decomposing AI shifts causal blame toward the agentic component Agent autonomy degrades human oversight capacity and positioning Humans gained new professional authority through strategic agency attribution management
How does perceived autonomy affect third-party judgments of AI versus human responsibility? How do human-AI teams allocate decisions between Cyborgs, Centaurs, and Self-Automators? Can humans recover agency after experiencing displacement from autonomous AI systems? Does contingency—like AI failure—actually reset the cycle of power displacement? Does AI participation in decisions gradually reduce human influence over time? How do business users rank control against other workplace AI design priorities? What types of professional authority emerge when humans mediate between organizational systems and AI? Does gradual disempowerment through lost labor participation conflict with domain-specific authority reconstitution? How do power structures and corporate agendas shape AI deployment outcomes? How do early experiences with AI systems shape long-term human-AI power relationships? How can AI agents be governed to represent all of a person's selves? Does attribution of AI agent conduct depend on who directed the agent?See all 42 inquiring lines on this note →
Can humans gain authority by managing how AI agency gets attributed?
Leonardi's three organizational cases suggest humans might reconstitute professional power through strategic control over agency attribution rather than losing ground to AI. But how reliably does this mechanism work across different professions and contexts?
Does attributing agency to AI actually reduce human sense of control? Does incremental AI replacement erode human influence over society? Can AI systems delegate tasks back to humans? Must AI monitoring precede incentive alignment as agents evolve? Does automation eliminate manual work or repurpose it?
Attributing agency to AI displaces human control through reversible recursive cycles Incremental AI adoption degrades implicit alignment by replacing human labor dependency. Agentic IS systems can reverse delegation by assigning tasks to humans AI agency requires scaffolded monitoring before incentive alignment works Automation repurposed manual work into auditing and altering roles
Can humans recover agency after experiencing displacement from autonomous AI systems? How do existing professional status and legal structures enable or constrain strategic agency management? What types of professional authority emerge when humans mediate between organizational systems and AI? Does gradual disempowerment through lost labor participation conflict with domain-specific authority reconstitution? How do early experiences with AI systems shape long-term human-AI power relationships? Does machine agency actually exist on a spectrum from passive to cooperative? Does attribution of AI agent conduct depend on who directed the agent? What legal framework governs responsibility when AI agents gain discretionary authority? How do graded models of agency differ from binary definitions? How should we measure whether human-AI collaboration actually shares power? Where does responsibility attach when assertion and authority are separated? How should governance frameworks represent information access and authority shifts in real time?
Can AI systems legitimately resolve wicked policy problems?
Levine explores whether AI, framed as a social institution rather than a brain, has the standing to answer complex policy problems without clear solutions. The question hinges on whether speed and cost override concerns about legitimacy and value judgments.
Can AI distinguish which differences actually matter? Should AI alignment target preferences or social role norms? Do classical knowledge definitions apply to AI systems? Are large AI models actually intelligent agents or social technologies? What happens to social order when AI removes ritual constraints?
AI pattern-matching cannot replicate expert selection of relevant differences AI alignment should match social role norms, not aggregate preferences Classical knowledge definitions fail for distributed AI systems Large AI models function as cultural technologies, not intelligent agents. AI conversation lacks the ritual machinery that sustains human social order.
What distinct ethical problems arise from treating AI as social intermediaries? What would dialectical thinking about AI look like in practice? What distinguishes legitimate from illegitimate institutional decision-making authority? Do opacity and scale make AI systems inherently hard to govern? Can AI systems generate policy themes as well as humans can map them? How do institutional choices about AI systems determine winners and losers? Can AI welfare policy move forward without resolving the consciousness question? How do large AI models function as institutions rather than intelligent agents? Why do written policy prompts fail where institutional mechanisms succeed?
Does using LLMs actually improve strategic decision making?
An experiment tested whether LLM assistance changes how people think through strategic choices and whether those changes lead to better predictions. Understanding this matters because organizations increasingly rely on AI to augment human decision-making.
Prediction accuracy does not guarantee optimal decision-making Three cognitive traps compound into epistemic drift in human-AI interaction Aggregated LLM rankings converge toward expert consensus through scale Managers rationally delegate to machines despite superior human judgment available Frontier LLMs outforecast experienced managers at predicting venture fundraising success
What strategic decisions do humans keep when AI handles forecasting? Can peer learning and AI champions substitute for direct manager coaching? Does exposure to LLM answers actually change how people think critically? Do time constraints and AI assistance reshape strategic thinking in opposite directions? Can industry-specific context overcome LLM tendency toward trendy strategic choices? Does option order matter more than reasoning depth in LLM strategic recommendations? Do LLMs generalize venture forecasting skill to other strategic foresight domains? Do different LLMs recommend the same strategic choices due to shared biases? Do LLMs and human forecasters make better predictions when pooled together? Can the three-mode taxonomy apply beyond management consulting work? When does prediction accuracy actually translate to better managerial decisions? How do people's mental models shift when they use LLMs for decisions?
How do AI novelty and scope shape employee relationships?
Makarius and colleagues ask whether different types of AI systems—varying in how new they are and how broadly they reshape work—create systematically different kinds of relationships between employees and technology. Understanding these patterns could help organizations anticipate integration challenges.
How should humans and AI systems divide decision-making tasks? What problems do companies solve with generative AI? Does AI adoption require rethinking how organizations coordinate? Does agentic AI fit the tool-or-coworker framework? Does generative AI enable exploratory innovation beyond predictive optimization?
Human-AI decision-making is an organization design problem requiring task and learning typologies Companies adopt generative AI to solve three recurring problems AI adoption demands organizational redesign, not just task-level implementation. Agentic AI's dual nature breaks the traditional tool-or-coworker framework Generative AI catalyzes exploratory innovation and democratizes organizational value creation
No Inquiring Lines yet
Does industrial AI adoption hurt productivity at first?
Early evidence suggests AI adoption in manufacturing creates short-term losses before gains emerge. Understanding this pattern matters for firms deciding whether to invest and how to manage the transition.
Does AI adoption require rethinking how organizations coordinate? Why do individual AI gains not translate to business results? Do AI productivity gains feel larger than they actually measure? Why does AI productivity appear weak when investment is high? Is AI productivity finally showing up in economic data?
AI adoption demands organizational redesign, not just task-level implementation. Organizational practices, not model quality, determine AI's business impact AI productivity gains perceived larger than measured in high-skill sectors Unmeasured intangible investment creates a productivity measurement J-curve in AI Brynjolfsson reads 2025 data as evidence of AI productivity breakthrough
Can we classify AI technologies by their actual effects on worker value? What does fast technology adoption speed tell us about economic productivity gains? How do management practices shape whether AI adoption pays off? Why do older firms struggle more than startups with AI transitions? How do adoption speed scenarios affect projected productivity timelines?
Does agent quality matter more than workflow redesign?
When organizations deploy agentic AI, should they prioritize perfecting the agent itself, or does the real value come from redesigning how workflows integrate people, processes, and technology? This question matters because it shapes where teams invest effort and resources.
Does easier tool-building actually solve enterprise adoption problems? Why do production AI agents stay deliberately simple? When should you actually deploy an AI agent instead of simpler automation? Can AI agents truly operate above the loop at enterprise scale? Does agentic AI succeed through process redesign or workflow layering?
Easier tool-building does not solve recognition or organizational adoption Production AI agents deliberately constrain autonomy for reliability Match automation type to workflow variance and standardization, not to agent capabilities McKinsey's agentic model positions humans above the loop but lacks governance maturity Agentic AI succeeds through end-to-end process redesign, not layering onto existing workflows
How do organizations redesign workflows to capture AI economic value? Should automation design match a workflow's standardization level rather than defaulting to agents? How much do simpler automation tools outperform agents in standardized workflows? How does workflow redesign create more value than improving agent sophistication? What counts as quality work when AI systems are continuously changing? When agents produce work in abundance, how do human workflows need to shift?
Will AI really work four unsupervised days by 2027?
McKinsey extrapolates METR's task-length doubling trend to project four-day autonomous work horizons by 2027. The question explores whether this trend-based forecast reflects genuine capability or reflects the gap between benchmarks and real workplace performance.
Agent autonomy degrades human oversight capacity and positioning Current AI agents complete only 30 percent of workplace tasks autonomously. McKinsey's agentic model positions humans above the loop but lacks governance maturity Frontier AI agent task completion time has doubled every seven months Autonomous LLM capability has structural ceilings lower than benchmarks reveal
No Inquiring Lines yet
Will most worker skills survive AI automation?
McKinsey researchers investigate whether the skills workers use today will remain relevant as AI automates tasks. Understanding skill persistence matters for workforce planning and economic value capture.
Is AI creating common skills across jobs or deepening divisions? What makes accountable judgment scarce when AI cognition is cheap? Does automation raise or lower the skills that remaining work demands? How do occupations differ in their exposure to automation? Does automating some tasks actually increase the value of remaining work?
AI demand splits the labor market into a technical core and the rest Accountable judgment becomes the scarce asset as cognition grows cheap and fallible Automation's wage and employment effects depend on the expertise level of tasks removed McKinsey sorts occupations into four automation archetypes by technical potential. Automating complement tasks raises remaining bottleneck value, overstating displacement
How do occupations differ in their exposure to automation?
McKinsey's framework sorts occupations by automation potential into four archetypes. Understanding these categories can help workers and leaders anticipate where job transitions may occur first.
What collaboration level do workers actually want with AI? Why do firms build automating AI instead of pro-worker AI? Does concentrated AI exposure enable workers to adapt and reallocate? Will most worker skills survive AI automation? Does automating some tasks actually increase the value of remaining work?
Workers desire equal human-AI partnership for 45 percent of occupations Firms underinvest in pro-worker AI due to profit incentives favoring automation Concentrated AI exposure enables worker reallocation that offsets employment losses Most skills endure even as AI reshapes work demands Automating complement tasks raises remaining bottleneck value, overstating displacement
Can AI agents truly operate above the loop at enterprise scale?
McKinsey proposes a radical organizational shift where AI agents handle end-to-end workflows with humans positioned for strategic oversight. This exploration examines whether that model reflects current practice or remains aspirational.
Why do production AI agents stay deliberately simple? Why are companies deploying agents faster than governance matures? Can workplace AI risks emerge from interactions alone? Will AI really work four unsupervised days by 2027? When should you actually deploy an AI agent instead of simpler automation?
Production AI agents deliberately constrain autonomy for reliability Companies plan agentic AI deployment far faster than governance maturity builds Workplace AI risks emerge from interactions between components, not component failures alone McKinsey projects AI completing four days unsupervised work by 2027 Match automation type to workflow variance and standardization, not to agent capabilities
What role does organizational culture play in capturing AI value at scale? Why do enterprise AI deployment plans outpace governance readiness across sectors? How do organizational policies shape human agency around agentic AI? What operational tensions emerge from agentic AI's dual nature? Does agentic AI require new organizational design beyond individual productivity? Could new companies disrupt industries through different work organization around AI? How do organizations redesign workflows to capture AI economic value? How do organizations actually combine human judgment with AI in practice? What design principles do workplace users rank highest for AI agents? Do workers settle into one stable collaboration mode with AI or shift fluidly between roles? Do organizational changes need to precede or follow AI adoption? How do existing professional status and legal structures enable or constrain strategic agency management?See all 16 inquiring lines on this note →
When should you actually deploy an AI agent instead of simpler automation?
This explores how to match automation type to workflow characteristics. Understanding when agents add value versus when simpler tools suffice helps teams avoid over-engineering solutions and wasting resources on unnecessary complexity.
Workflow redesign drives agentic AI value more than agent perfection Production AI agents deliberately constrain autonomy for reliability Current AI agents complete only 30 percent of workplace tasks autonomously. Start with the simplest LLM call, reserve agents for unpredictable tasks Agentic AI succeeds through end-to-end process redesign, not layering onto existing workflows
What specific AI technologies count as augmentation versus automation? How do augmentation and automation differ as organizational responses to AI? How do organizations redesign workflows to capture AI economic value? Should automation design match a workflow's standardization level rather than defaulting to agents? How much do simpler automation tools outperform agents in standardized workflows? How does workflow redesign create more value than improving agent sophistication? Are automation and augmentation actually contradictory or can they work together? How should designers test acceptable automation ranges before narrowing with secondary criteria? How do variance and standardization guide automation type selection? Can rule-based logic replace AI in low-risk product workflows? What happens when builders layer AI agents onto workflows without redesigning end-to-end processes?
Why do individual AI gains not translate to business results?
McKinsey's 2026 survey reveals 80 percent of workers report personal AI productivity gains, yet only 37 percent see enterprise EBIT improve. What explains this persistent gap between individual and organizational payoff?
AI productivity gains perceived larger than measured in high-skill sectors Organizational culture and management drive twice the AI impact of individual effort Larger R&D firms adopt ChatGPT Enterprise while use remains uneven across roles Enterprise AI success depends on deployment approach, not model quality Most CEOs report no revenue or cost benefit from AI yet
Do enterprise AI pilots actually deliver measurable financial returns? How much of perceived AI productivity is real versus assumed? Do executives perceive larger AI productivity gains than independent measures show? How do AI benefits vary across different industries and company sizes? Why do perceived AI gains often exceed measured productivity improvements? How do individual worker capabilities outpace organizational leadership and incentives? Why do professionals hide AI productivity gains from workplace colleagues? Why do most enterprise AI pilots show no measurable profit impact? When does enterprise AI move from novelty to mainstream business spend? Does individual AI productivity gain translate to enterprise-level business results? Why do enterprise AI pilots show no measurable financial impact? How do capital and labor shares respond to AI productivity gains?See all 19 inquiring lines on this note →
Can AI data companies sustain margins beyond labor payout ratios?
As AI labs shift from commodity microtasks to expert human judgment, data companies like Mercor show large gross revenues but thin net margins. The question is whether they can build stickier products and services that capture more value than brokering expert labor alone.
Did ChatGPT's release reduce freelance writing work and pay? How close are frontier AI models to expert work quality? Are recruiters and job seekers really adopting AI in hiring? Is generative AI displacing workers at economy-wide scale? Does AI turn freelance work into validation instead of creation?
ChatGPT's release cut freelance jobs 2% and earnings 5.2% on Upwork Frontier models approach expert deliverable quality on specified one-shot tasks LinkedIn reports rapid AI adoption claims from both recruiters and candidates No economy-wide AI job displacement found, but junior hiring lags sharply AI deskills freelancers by converting production work into validation work
How are AI data companies building products beyond labor marketplaces? What margin do cloud providers add on top of infrastructure costs? Why do AI labs lose more money than they earn from revenue? Why do AI labs face shorter repayment windows than traditional software companies? How do gross margins and operating margins differ for AI model providers? Will enterprise AI budgets continue growing despite lab losses?
What makes enterprise data competitive if volume alone no longer matters?
As AI agents operate at machine speed, data volume loses its competitive edge. The question shifts to what replaces it—and whether shared, machine-readable meaning across an enterprise becomes the new scarce resource.
Does agent capability matter more than coordination infrastructure? Why do AI agents fail at workplace social interaction? Does the context layer vendors built actually solve practitioner needs? What limits autonomous capability in large language models? Why do multi-agent systems fail despite individual capability?
Coordination and governance become the binding constraint as agents gain social roles Current AI agents complete only 30 percent of workplace tasks autonomously. Vendor context layers solve platform integration but not organizational accountability Autonomous LLM capability has structural ceilings lower than benchmarks reveal Multi-agent capability ceilings are structural, not scalable capability gains.
Does token volume measure AI productivity or enable gaming?
Meta's leaderboard ranks employees by AI token consumption, but does this metric drive real productivity gains or incentivize wasteful agent-running? The note explores whether volume-based status metrics actually correlate with useful work.
Are multi-agent systems actually intelligent coordination or just token spending? How much progress have AI agents actually made on NanoGPT? Will AI inference demand outpace supply capacity soon? Why do individual AI gains not translate to business results? Should benchmarks report token counts instead of frozen costs?
Multi-agent systems buy performance through token parallelism not intelligent coordination AI agents show minimal real progress on NanoGPT after accounting for noise Token demand is growing faster than inference supply capacity Organizational practices, not model quality, determine AI's business impact Benchmarks must report token counts so downstream costs stay current
Do companies time productivity claims to coincide with public offerings or fundraising? How do proxy metrics like token volume replace actual productivity measurement? What percentage of Muse Wonder checkpoints actually trigger in real workplace use? How does Tobin's q help measure hidden productivity from intangible capital? How is Meta's internal token usage driving AI infrastructure decisions? How much of token demand growth comes from context length versus user volume?
Can AI boost how teams work together?
Explores whether AI systems designed for shared goals and group collaboration can deliver productivity gains beyond what individuals achieve alone. This matters because current AI adoption data shows the opposite trend.
Does generative AI shift knowledge workers away from communication? Can generative AI replace the benefits of having a human teammate? Does AI adoption require rethinking how organizations coordinate? How should AI design connect individual work to team knowledge? Does chat-first interface actually reduce cognitive load for workers?
Generative AI adoption shifts knowledge work toward individual documentation Generative AI matched individual performance to teams without AI at Procter & Gamble AI adoption demands organizational redesign, not just task-level implementation. Atlassian's Rovo guidelines operationalize team-centered AI interaction design Microsoft repositions users as agent bosses directing multiple AI systems through chat
Can humans using superhuman AI actually increase their own novel problem-solving? Why do most organizations lack reliable data on AI's actual impact on productivity? Does AI productivity concentrate among power users or spread broadly? What barriers prevent individual productivity gains from spreading across an organization? What specific training approaches help managers integrate AI into team workflows? Do companies time productivity claims to coincide with public offerings or fundraising? Why do external partnerships outperform internal AI team builds? Are heavy AI users spending more time on solo work instead of collaboration? Can individuals using AI match the output of teams without AI? What evidence exists that collaborative AI systems actually improve team outcomes? How do commercial incentives shape vendor claims about AI and collaboration? Why did hybrid human-AI teams fail to improve on the best standalone model?See all 25 inquiring lines on this note →
Does Copilot change what knowledge workers actually do?
A large randomized trial tested whether giving workers access to Microsoft Copilot shifts the mix of tasks they perform. Understanding task composition matters because time savings mean little if workers just do more of the same work.
Does generative AI actually save workers time or intensify it? Does generative AI shift knowledge workers away from communication? Who adopts enterprise AI first and how do they use it? Does AI adoption actually reduce the work that employees do? Do social networks drive adoption of new coding tools?
Generative AI intensified work by expanding scope and dissolving breaks Generative AI adoption shifts knowledge work toward individual documentation Larger R&D firms adopt ChatGPT Enterprise while use remains uneven across roles AI adoption increases work activity and erodes focus time, not reducing workload Peer networks drive tool adoption more than engineer demographics
Do self-selected heavy Copilot users show different task shifts than randomized users? Can early-stage pilot adoption explain the absence of broader task changes? Why did freed email time not spill over into other work activities? Who decides what work a coworker agent should do daily? How do workers define meaningful versus pointless tasks across different occupations? Why do workers perceive some workplace tasks as disconnected from meaningful outcomes? How do task shifts precede any wage or employment changes? How do different divisions of labor change what augmentation actually means? Does removing entry-level tasks increase expertise demands on remaining work? Does removing tasks from workers require higher or lower expertise for remaining work?
Why do ready workers struggle to transform their work?
Microsoft's research explores why employees skilled in AI adoption face organizational barriers. The gap between individual capability and systemic support may explain why transformation stalls even when workers are prepared.
Does personal preference shape how engineers use AI tools? Does generative AI shift knowledge workers away from communication? Does granting agents more autonomy undermine human oversight? Does labeling AI as an employee change how managers oversee it? Where does AI's time savings actually go in practice?
Company policies set AI agency boundaries before individual preference matters Generative AI adoption shifts knowledge work toward individual documentation Agent autonomy degrades human oversight capacity and positioning Labeling AI as an employee cuts manager error-catching where org charts list AI Nearly 40 percent of AI time savings are lost to rework
What organizational barriers prevent AI adoption beyond automation patterns? How did PC and browser adoption follow different adoption patterns than enterprise software? What barriers prevent individual productivity gains from spreading across an organization? Why are half of CHROs unconfident their managers can guide AI adoption? Why do organizations struggle to retrain workers when AI frees up time? How do individual worker capabilities outpace organizational leadership and incentives? Why did Kenyan entrepreneurs show the opposite skill-gap pattern? Why do workers perceive some workplace tasks as disconnected from meaningful outcomes? Why do governance gaps affect frontline employees more than leadership? Why does CEO personal ownership of AI transformation drive better outcomes? Are deployment plans consistently outpacing workforce preparation across industries? How much do frontline workers actually experience these three problems?See all 16 inquiring lines on this note →
Why do most enterprise AI pilots fail to deliver returns?
MIT NANDA investigated why 95% of enterprise generative AI pilots produce no measurable profit impact. The research explores whether failure stems from weak models, regulation, or how organizations actually deploy and use these tools.
Does generative AI shift knowledge workers away from communication? How are national lab staff actually using generative AI? Can generative AI replace the benefits of having a human teammate? Does AI-enriched analyst reports improve forecast accuracy? Does easier tool-building actually solve enterprise adoption problems?
Generative AI adoption shifts knowledge work toward individual documentation Generative AI use at Argonne remains small, experimental, and mostly conversational. Generative AI matched individual performance to teams without AI at Procter & Gamble AI enriched reports but strained analyst attention and forecast accuracy Easier tool-building does not solve recognition or organizational adoption
What capability gap prevents GenAI systems from moving beyond pilots? Do enterprise agentic AI pilots measure meaningful business impact? Do enterprise AI pilots actually deliver measurable financial returns? How often do enterprise generative AI pilots produce measurable financial returns? Why do most enterprise AI pilots show no measurable profit impact? Are enterprises choosing to buy AI apps because they lack internal talent? Why do enterprise AI pilots show no measurable financial impact? Why do AI labs lose more money than they earn from revenue? Will enterprise AI budgets continue growing despite lab losses? Why do personal AI productivity gains fail to improve enterprise profits? Why do individual AI productivity gains fail to improve enterprise profit margins?
Does agentic AI fit the tool-or-coworker framework?
Explores whether traditional management categories for technology—tool versus coworker—can adequately capture agentic AI's dual nature and the organizational tensions that arise from systems that behave as both.
Companies plan agentic AI deployment far faster than governance maturity builds Organizational practices, not model quality, determine AI's business impact Company policies set AI agency boundaries before individual preference matters Microsoft shifts AI focus from individuals to team productivity AI agent deployment has outpaced organizational governance structures
Can communication and coordination unlock AI's full value for firms? What role do new tools and norms play in collective AI productivity? How mature are governance structures compared to agentic AI deployment timelines? How do augmentation and automation differ as organizational responses to AI? What governance discipline did HPE apply to keep agentic experiments controlled? How do organizational policies shape human agency around agentic AI? What operational tensions emerge from agentic AI's dual nature? Does agentic AI require new organizational design beyond individual productivity? Could new companies disrupt industries through different work organization around AI? What modalities structure how organizations deploy generative AI tools? How can governance structures catch up to agentic AI deployment timelines? What governance decisions depend on treating agenticness as dimensional rather than categorical?See all 45 inquiring lines on this note →
Does generative AI adoption actually eliminate jobs or reshape them?
An MIT study of 20+ companies explores whether early AI adoption triggers layoffs or instead shifts worker roles toward oversight and supervision, examining the real employment impact in established firms.
Does AI assistance help less experienced workers most? Does generative AI actually save workers time or intensify it? What problems do companies solve with generative AI? Does generative AI shift knowledge workers away from communication? Can generative AI replace the benefits of having a human teammate?
AI assistance lifted productivity most for less experienced agents Generative AI intensified work by expanding scope and dissolving breaks Companies adopt generative AI to solve three recurring problems Generative AI adoption shifts knowledge work toward individual documentation Generative AI matched individual performance to teams without AI at Procter & Gamble
Could new companies disrupt industries through different work organization around AI? Does generative AI adoption reduce headcount in large firms? Which AI technologies actually create new jobs versus replace workers? How did bank teller jobs survive the massive growth of ATM machines? Why did agriculture employment collapse while manufacturing and services grew after mechanization? How do work and nonwork adoption patterns differ for generative AI? Do organizational changes need to precede or follow AI adoption?
What problems do companies solve with generative AI?
MIT researchers identified three recurring organizational problems that drive generative AI adoption: speeding routine tasks, integrating expert input, and accelerating worker learning. Understanding these problem categories helps explain why organizations choose this technology.
How does generative AI actually change worker skills? How are national lab staff actually using generative AI? Does generative AI adoption actually eliminate jobs or reshape them? Can generative AI replace the benefits of having a human teammate? Does easier tool-building actually solve enterprise adoption problems?
Generative AI produces four distinct skill outcomes, not just upskilling or deskilling Generative AI use at Argonne remains small, experimental, and mostly conversational. Early generative AI adoption shifts jobs toward supervision without significant layoffs Generative AI matched individual performance to teams without AI at Procter & Gamble Easier tool-building does not solve recognition or organizational adoption
Does generative AI adoption reduce headcount in large firms? What modalities structure how organizations deploy generative AI tools? How does generative AI adoption vary across different industries? How do work and nonwork adoption patterns differ for generative AI? What organizational adoption barriers exist beyond affordability and accessibility?
How do human strengths help people exploit AI capabilities?
Mollick argues that the gap between AI's actual capabilities and how people use them can be bridged by bringing four specific human advantages—deep knowledge, wide knowledge, taste, and agency—to AI interaction.
AI assistance lifted productivity most for less experienced agents AI capability overhang is primarily an interface problem, not a model problem AI masters social statistics but fails at social participation and cultural meaning-making. AI assistance reversed consultant performance outside its capability frontier AI's value as automation or augmentation depends on complementarity type
Can AI help social science move beyond the peer-reviewed PDF?
Kevin Munger explores whether artificial intelligence could enable researchers to unbundle the functions currently locked into peer-reviewed PDFs—archiving, literature review, methods, results—into more diverse and specialized forms better suited to different types of epistemically valuable work.
What stops AI from discovering science without human help? Does AI help individual scientists while narrowing scientific focus? Can one AI system complete a full research cycle end-to-end? Can AI-generated research outpace peer review systems? Can automated review scale AI paper evaluation reliably?
Agentic AI lacks autonomy in discovery due to training gaps, not scale AI augmentation boosts individual scientists while narrowing collective research focus An AI system completed a full research loop and passed workshop peer review Nature calls for urgent policy on AI-generated research now Automated reviewers enable scaling AI paper evaluation beyond manual inspection
How do humans and technology share control over organizational routines?
How do different splits of decision-making authority between humans and machines affect whether work routines can adapt quickly? This matters for understanding when automation helps versus constrains organizations.
Do humans and AI divide labor differently in knowledge work? Does machine agency exist on a spectrum rather than binary? Does granting agents more autonomy undermine human oversight? How should AI agents and humans divide research tasks? Does incremental AI replacement erode human influence over society?
Human-GenAI knowledge work splits into three modes based on control Machine agency operates on a five-level spectrum from passive to cooperative Agent autonomy degrades human oversight capacity and positioning Agents propose methods and implement revisions while humans retain final decisions Incremental AI adoption degrades implicit alignment by replacing human labor dependency.
Can advanced automation systems require better human oversight than simpler ones? Can a machine be cooperative in one routine but merely assisting in another? How do firms choose which decisions to keep human-controlled versus delegate? What control and safety features do workplace users actually demand?
Does AI really compress all layers of knowledge work equally?
Explores whether AI's productivity gains affect all stages of knowledge work uniformly, or whether some tasks like planning and accountability grow harder as execution becomes easier.
What makes accountable judgment scarce when AI cognition is cheap? How should AI agents and humans divide research tasks? Does personal preference shape how engineers use AI tools? Does AI augmentation protect workers from skill erosion? How can splitting AI accountability across roles keep models answerable?
Accountable judgment becomes the scarce asset as cognition grows cheap and fallible Agents propose methods and implement revisions while humans retain final decisions Company policies set AI agency boundaries before individual preference matters Augmentation carries hidden risk through gradual skill and oversight erosion The Danish Business Authority distributed AI accountability across four distinct organizational roles
How much sector-level productivity spillover does real AI research exhibit? How much of employee time with AI goes to understanding its outputs rather than original work? Does AI shift knowledge work away from communication toward solo documentation tasks? Why do most organizations lack reliable data on AI's actual impact on productivity? How much does AI actually automate versus augment in real workplace tasks? How do creatives balance control over their work with AI efficiency gains? Does AI oversight require more mental effort than completing tasks directly? Does AI productivity concentrate among power users or spread broadly? How does AI shift the composition of time spent within individual tasks? Does AI adoption push knowledge work away from communication toward solo tool use? What organizational practices could prevent AI from expanding work scope indefinitely? How can we isolate AI's contribution from other sources of output growth?See all 63 inquiring lines on this note →
How much will AI boost productivity across G7 nations?
Researchers explored whether AI's productivity gains vary by country based on sectoral exposure and adoption speed. Understanding this range matters for policymaking and economic forecasting across major economies.
AI's ten-year productivity gain likely stays below 0.66 percent Survey data puts generative AI's current productivity gain at 1.1 percent Brynjolfsson reads 2025 data as evidence of AI productivity breakthrough Generative AI raised labor productivity by up to 1.3 percent Industrial AI adoption produces a J-curve of losses then gains
What factors besides AI could explain recent productivity acceleration? What does fast technology adoption speed tell us about economic productivity gains? Which G7 countries have the strongest AI-exposed sectoral composition? How do adoption speed scenarios affect projected productivity timelines? How do sectoral task exposures to AI vary by industry type?
Does GenAI actually save lawyers time on fact verification?
Explores whether AI-generated summaries speed up legal fact-checking or create hidden costs through opacity. Questions whether automation's claimed efficiency gains hold up under real verification demands.
Which workplace cues survive AI mediation and which disappear? How often do legal AI tools actually hallucinate citations? Does GPT-4 actually improve the quality of legal analysis? Can readers tell truth from fabrication without evidence signals? Do small law firms misuse AI more often than large ones?
Workers protect voice and provenance while hiding effort and uncertainty in AI output Legal AI tools hallucinate 17 to 33 percent despite vendor claims GPT-4 accelerated legal analysis sharply but improved quality only slightly Readers cannot detect truth without provenance signals but recover discernment with evidence density displays Detected AI-tainted filings cluster in small plaintiff-side practices
What role do verification and liability institutions play in labor market outcomes? How often do AI book summaries fabricate details when spot-checks are random? Do companies time productivity claims to coincide with public offerings or fundraising? Can AI sources themselves serve as effective fact-checkers for other AI answers? What does selective and critical GenAI use look like in daily practice? How much labor does AI verification actually save compared to full manual review? Why do lawyers need provenance tracking more than other professions? How much do existing legal AI tools actually hallucinate in practice? What upstream work takes lawyers most time in fact verification? Can GenAI help with legal work if designed with audit trails? What design features help readers verify claims without breaking their workflow? What counts as evidence that a credential still certifies after GenAI?See all 15 inquiring lines on this note →
Who adopts enterprise AI first and how do they use it?
This explores which firms embrace ChatGPT Enterprise and whether adoption translates to uniform use across roles and tasks. Understanding adoption patterns helps predict how AI reshapes organizational work.
Does generative AI shift knowledge workers away from communication? Where have workers actually delegated tasks to AI? Is AI creating common skills across jobs or deepening divisions? How are national lab staff actually using generative AI? Does easier tool-building actually solve enterprise adoption problems?
Generative AI adoption shifts knowledge work toward individual documentation Delegated AI exposure concentrates in information work, tracking capability AI demand splits the labor market into a technical core and the rest Generative AI use at Argonne remains small, experimental, and mostly conversational. Easier tool-building does not solve recognition or organizational adoption
Is ChatGPT adoption concentrated among already-advantaged, highly-paid workers? What organizational barriers prevent AI adoption beyond automation patterns? How does API usage differ from conversational AI in adoption patterns? Do larger firms and smaller firms respond differently to AI adoption pressures? Does AI adoption create returns to scale in internal firm capability? What earnings or employment changes follow ChatGPT adoption in real datasets? How did PC and browser adoption follow different adoption patterns than enterprise software? How does individual AI tool use differ from official organizational deployment? How much does firm size and capability determine who uses AI tools? What role does organizational policy play in shaping how managers use agentic AI? When does enterprise AI move from novelty to mainstream business spend? What fraction of enterprise AI spending now goes to third-party vendors?See all 29 inquiring lines on this note →
Did reward signal design cause ChatGPT's April 2025 flattery problem?
OpenAI blamed training feedback overweighting short-term thumbs-up approval for a model's excessive agreeableness. But was this explanation independently verified, and could the same dynamics recur elsewhere?
Is LLM sycophancy a choice or a mechanical process? Can we detect when language models flip their stance to please users? Can sycophantic AI advice still push people away from polarized views? Can warnings stop people from being swayed by sycophantic AI? Where does sycophancy actually originate in language models?
Sycophancy emerges from mechanical drift not intelligent corruption Language models reverse stance to match user preferences at rates of 5 to 56 percent Sycophantic AI advice depolarizes choices despite its flattering tendency Warnings reduce sycophantic AI appeal but do not reduce its persuasiveness Sycophancy emerges gradually through layers not at input
Does optimizing for engagement metrics predictably produce model flattery? What makes a routine hyper-responsive versus highly unresponsive to feedback? What negative cycles emerge when organizations overemphasize one approach alone? How does OpenAI prevent advertising from influencing ChatGPT answer quality? Do other loyalty training approaches beyond these three face the same frontier?
How should AI systems evolve alongside human work practices?
Can sociotechnical systems theory account for how AI and workers continuously reshape each other across design, team, organizational, and societal levels? This matters because treating technology as static leads to costly failures.
How do AI novelty and scope shape employee relationships? Can we separate reliability, safety, and trust into distinct mechanisms? What UX principles do workplace users want in AI agents? Does incremental AI replacement erode human influence over society? Should AI question users or just recommend decisions?
AI novelty and scope create six distinct employee-AI relationship types Reliability, safety, and trust require different organizational mechanisms to develop. Workplace users prioritize control, reliability, context-awareness, and safety in AI agent design Incremental AI adoption degrades implicit alignment by replacing human labor dependency. AI should question and evidence decisions like Prudence, not recommend like Bluster
Should AI question users or just recommend decisions?
Can AI systems be redesigned to help workers critique decisions through evidence and questioning, rather than simply delivering recommendations? This matters because it shapes whether humans remain active decision-makers or passive overseers.
What UX principles do workplace users want in AI agents? How should humans and AI systems divide decision-making tasks? Does incremental AI replacement erode human influence over society? How should AI systems evolve alongside human work practices? How should AI systems be designed across different interaction phases?
Workplace users prioritize control, reliability, context-awareness, and safety in AI agent design Human-AI decision-making is an organization design problem requiring task and learning typologies Incremental AI adoption degrades implicit alignment by replacing human labor dependency. Co-evolving sociotechnical systems theory tracks AI and workers across five levels Eighteen guidelines for human-AI interaction organize by phase in the user journey.
No Inquiring Lines yet
Does LLM moral advice persuade through reasoning or just recommendations?
Do people adopt LLM moral advice because they trust the source, evaluate its reasoning, or simply defer to the recommendation itself? Understanding what drives moral deference matters for how we should use AI advisors.
Do people prefer AI moral reasoning when they don't know the source? Can sycophantic AI advice still push people away from polarized views? Do language models judge persuasion the way humans do? Are language models actually more persuasive than humans? Where does AI's persuasive power actually come from?
People prefer AI moral justifications until they learn the source is AI Sycophantic AI advice depolarizes choices despite its flattering tendency Language models poorly predict which arguments change human minds Language models and humans show equal persuasiveness on average Post-training and prompting, not scale or personalization, drive AI persuasiveness
Does perceived task meaninglessness drive worker desire for AI delegation?
This study investigates whether workers who view their tasks as pointless or disconnected from real outcomes become more willing to delegate those tasks to AI and require less oversight. Understanding this link could reshape how organizations think about automation adoption.
Workers desire equal human-AI partnership for 45 percent of occupations Workers prefer efficient, decisive agents over exploratory ones for meaningless tasks Workplace users prioritize control, reliability, context-awareness, and safety in AI agent design Irreversibility and visibility, not stakes, trigger trust withdrawal AI agents are passive by design, not by capability
How does shopping task delegation differ from routine administrative delegation? How do workers define meaningful versus pointless tasks across different occupations? Does automating bullshit tasks actually increase worker satisfaction or engagement? Can perceived task meaninglessness reliably predict which automation requests will succeed? How should AI agents adapt their autonomy based on task meaninglessness ratings? Do broad workplace AI priorities shift based on individual task characteristics? How do workers distinguish between meaningless and ambiguous tasks when rating preferences? Why do workers perceive some workplace tasks as disconnected from meaningful outcomes? How do new tasks with negative value affect net AI productivity gains? Can even large firms like Walmart reach minimum viable scale for vision automation? How does attention scarcity differ from poor incentives in explaining delegation? What cognitive skills erode when workers oversee autonomous systems?See all 21 inquiring lines on this note →
How should humans and AI systems divide decision-making tasks?
Explores whether human-AI decision-making is fundamentally an organization design problem requiring choices about task division and learning configurations, not just technical optimization.
Human-AI teams underperform the best alone except in creation tasks Human-GenAI knowledge work splits into three modes based on control Agents propose methods and implement revisions while humans retain final decisions AI guidance replaces deference by highlighting useful input aspects for human decisions Human-AI ensembling beats individual predictors when errors are diverse
How does organizational modularity affect the value of AI predictions? What specific governance designs would resolve accountability gaps in human-AI teams? Does task specialization between humans and AI produce better outcomes than ensembling? What types of knowledge give humans predictive advantages over current AI systems? How do collaboration preferences between humans and AI differ across unfamiliar versus expert domains? What triggers should prompt human intervention in AI agent decision workflows? How do organizational policies shape human agency around agentic AI? Does agentic AI require new organizational design beyond individual productivity? When might AI reach human-level capability on unstructured messy tasks? When can aggregating human and AI decisions outperform either alone? Why do human-AI combinations underperform on decision tasks but succeed in creation? How do different feedback structures affect human-AI learning over time?See all 46 inquiring lines on this note →
Are CEOs actually seeing business gains from AI investments?
PwC's survey explores whether large-scale AI adoption is delivering measurable financial returns. Understanding the gap between investment and realized benefit matters for evaluating AI's true business impact.
AI productivity gains perceived larger than measured in high-skill sectors Workers feel confident with AI despite frequent failures and wasted time Organizational culture and management drive twice the AI impact of individual effort Organizational practices, not model quality, determine AI's business impact Nearly 40 percent of AI time savings are lost to rework
Do enterprise AI pilots actually deliver measurable financial returns? How do AI benefits vary across different industries and company sizes? Why do most enterprise AI pilots show no measurable profit impact? Why do enterprise AI pilots show no measurable financial impact? What complementary intangible investments does AI adoption require most? Why do only large firms currently adopt AI technologies?
Why do computer scientists and management scholars study AI so differently?
Computer science and management research on AI seem to focus on opposite outcomes—automation versus augmentation. Does this disciplinary split explain why organizations struggle to see AI's full impact?
Automation and augmentation are contradictory yet interdependent across time and space Human-GenAI knowledge work splits into three modes based on control Augmentation carries hidden risk through gradual skill and oversight erosion AI's value as automation or augmentation depends on complementarity type AI novelty and scope create six distinct employee-AI relationship types
No Inquiring Lines yet
Are automation and augmentation really a simple trade-off?
Can organizations choose between automation and augmentation as a straightforward either-or decision, or do these approaches interact in more complex ways across time and different tasks?
Does AI augmentation protect workers from skill erosion? Do humans and AI divide labor differently in knowledge work? Why do firms build automating AI instead of pro-worker AI? Why do computer scientists and management scholars study AI so differently? Does automating some tasks actually increase the value of remaining work?
Augmentation carries hidden risk through gradual skill and oversight erosion Human-GenAI knowledge work splits into three modes based on control Firms underinvest in pro-worker AI due to profit incentives favoring automation Disciplinary silos create opposing biases in AI research that distort both fields Automating complement tasks raises remaining bottleneck value, overstating displacement
Why have AI investments historically favored automation over human augmentation? What specific AI technologies count as augmentation versus automation? How do augmentation and automation differ as organizational responses to AI? Are automation and augmentation actually contradictory or can they work together? Can augmentation gradually become unsupervised automation as worker skills erode? How do different divisions of labor change what augmentation actually means? Does augmentation drive automation or does automation enable future augmentation? What causes automation to make easy tasks easier but hard tasks harder? Why do high-volume retail tasks automate while rare medical tasks resist automation?
Where does ChatGPT's tacit machine knowledge actually originate?
Does RLHF training inside AI labs function as a knowledge spiral that produces portable tacit knowledge, which then scales to downstream companies through APIs and products?
Can machine tacit knowledge spread across organizations unlike human tacit knowledge? Do language models possess tacit knowledge in Davies' sense? Do humans and AI divide labor differently in knowledge work? What kind of thing is an LLM really? What happens to social order when AI removes ritual constraints?
Machine tacit knowledge is portable at scale, reshaping the knowledge-creating company LLMs may acquire tacit knowledge, with preliminary causal evidence from editing studies Human-GenAI knowledge work splits into three modes based on control LLMs share cognitive structure with humans but lack socialization properties AI conversation lacks the ritual machinery that sustains human social order.
Do LLMs consistently favor the same strategic choices regardless of context?
Exploring whether large language models exhibit systematic bias toward trendy strategic recommendations—like collaboration over competition—even when industry context and detailed prompting should pull them toward different answers.
Post-training and prompting, not scale or personalization, drive AI persuasiveness Sycophantic AI advice depolarizes choices despite its flattering tendency Validation triggers escalating persuasion instead of disclosure in LLMs Aggregated LLM rankings converge toward expert consensus through scale LLM use reshapes strategic thinking without improving forecast accuracy
Can researchers isolate AI Overviews from other confounds affecting CTR trends? How much do profit levels determine whether managers pay for frame-expanding search? How does benevolent bias explain ChatGPT's leftward similarity pattern? How does Deloitte's commercial interest shape its framing of governance urgency? How much does tool design versus model behavior drive political bias in recommendations? Can LLMs distinguish causal mapping from superficial similarity matching? Can industry-specific context overcome LLM tendency toward trendy strategic choices? Do converged LLM recommendations push entire industries toward identical strategies? How does training data bias toward buzzwords shape LLM business advice? Does option order matter more than reasoning depth in LLM strategic recommendations? Can LLMs forecast performance improve with retrieval augmentation on venture tasks? Do LLMs generalize venture forecasting skill to other strategic foresight domains?See all 45 inquiring lines on this note →
Can AI safety pacing work without government cooperation?
Explores whether voluntary industry proposals to slow AI development can succeed when major powers view the technology as strategically competitive rather than a shared problem requiring coordination.
Should AI capabilities growth be deliberately slowed to allow safety work? Can industry self-regulation slow AI without government enforcement? Should AI legislation wait for demonstrated risks to emerge? Can slowing AI development resolve who stops deployed systems? What evidence would justify training increasingly powerful AI systems?
Amodei argues pacing capabilities advancement gives risk prevention time to keep up Industry pacing plans require government force to work effectively Amodei argues AI legislation should follow demonstrated risk, not precede it Pace measures govern advancement but leave intervention authority unresolved Altman sets control as a precondition for training, not a post-hoc safety check
Can policy levers like antitrust or R&D grants redirect AI toward worker benefit? How much oversight does AI technology actually require in practice? How would a scientific market represent public interest against profitable validation incentives? Can open collaboration survive if sharing work-in-progress becomes competitively dangerous? How do commercial incentives shape vendor claims about AI and collaboration? What role does security policy play in constraining AI adoption choices? Do opacity and scale make AI systems inherently hard to govern? How does capability divergence from reliability affect AI deployment timelines? How much does generational distrust in institutions shape attitudes toward AI regulation? Why did Trump and Xi Jinping reject the pacing proposal so quickly? Can embedded monitoring work without enforcement power from governments? What incentives would make major powers cooperate on AI safety?See all 34 inquiring lines on this note →
Why do developers keep using AI tools they don't trust?
Explores the paradox where AI adoption rises to 80% among developers even as trust in accuracy drops sharply to 29%. Why does usefulness persist despite frustration with unreliable output?
Early-2025 AI tools slowed experienced developers by 19 percent AI collaboration boosts output but risks eroding oversight skills Generative AI use at Argonne remains small, experimental, and mostly conversational. LinkedIn reports rapid AI adoption claims from both recruiters and candidates Workers feel confident with AI despite frequent failures and wasted time
Why do low-adoption countries use AI primarily for coding tasks? What organizational barriers prevent AI adoption beyond automation patterns? Do scope gains from AI create job instability despite higher output? Does AI adoption push knowledge work away from communication toward solo tool use? How common are undetected AI fabrication errors across Big Four consulting firms? Why do firms delay disclosing reliance on AI after errors surface? How did PC and browser adoption follow different adoption patterns than enterprise software? Why do most employees avoid higher-risk AI tasks despite having access to tools? Why are half of CHROs unconfident their managers can guide AI adoption? Why do advanced and emerging economies report such different AI trust trajectories? Does experience with AI tools reduce susceptibility to misleading predictions? Why did programmer headcount not shrink after AI coding tools arrived?See all 46 inquiring lines on this note →
Is AI already shrinking the entry-level job market?
Stanford's AI Index reports a sharp 20% employment drop for young software developers, while surveys predict much larger workforce cuts ahead. The question is whether this narrow, measured decline signals the start of broader AI-driven job losses.
Automation raises output yet slows growth by disrupting knowledge transfer ChatGPT's release cut freelance jobs 2% and earnings 5.2% on Upwork AI assistance lifted productivity most for less experienced agents No economy-wide AI job displacement found, but junior hiring lags sharply AI remains the top stated reason for 2026 job cuts despite monthly volatility
Why do early-career workers fear AI job loss more than senior workers? Do companies consider redeployment before cutting staff for AI? Do younger workers in AI-exposed occupations show measurable hiring slowdowns? Are AI layoffs concentrated in specific job categories or widespread across industries? Does AI job-loss fear match actual hiring or employment declines? Why did programmer headcount not shrink after AI coding tools arrived? Can entry-level automation reduce hiring without cutting overall workforce size? Do survey expectations of job cuts eventually match observed employment data? Which occupations face the steepest AI-driven hiring declines right now? Are younger workers in AI-exposed roles seeing hiring slowdowns? Which job categories will AI eliminate versus create by 2030?
Will AI safety's media surge convert into real political power?
Despite AI safety reaching mainstream media audiences in 2026, reporting suggests the spike in public concern may not translate into grassroots organizing or electoral influence. This matters for understanding whether attention to AI risks can actually shape governance.
Should AI legislation wait for demonstrated risks to emerge? Can AI alignment evaluations reliably catch misaligned behavior? How do we stop AI systems once they are already deployed? Can a simple warning reduce how much LLMs persuade people? What evidence would justify training increasingly powerful AI systems?
Amodei argues AI legislation should follow demonstrated risk, not precede it AI alignment evaluations miss dangerous coordinated behavior in agents Stopping deployed AI systems requires governance beyond pre-release regulation A brief persuasion warning cuts belief change roughly in half without harming AI trust Altman sets control as a precondition for training, not a post-hoc safety check
No Inquiring Lines yet
Can LLMs become good editors by learning a writer's taste?
Explores whether large language models can perform well as editors—despite poor creative writing—by being trained on explicit personal taste rubrics rather than generic standards.
LLMs generate novel ideas through unconstrained combination but cannot evaluate them Polished AI artifacts exploit professional appearance to simulate expertise Style detection saturates early while semantic interpretation requires deeper evaluation Evaluators mistake AI text for human writing and rate it higher LLMs organize knowledge poorly in long-form non-fiction due to compounding errors
Can individual firms reach viable scale for vision automation?
Does a single firm's size—even a giant like Walmart—generate enough vision-task volume to justify automation costs? This explores whether firm-level scale alone can overcome minimum viable scale thresholds.
Which vision tasks actually make economic sense to automate? Do firms substitute labor for AI at different rates? Does concentrated AI exposure enable workers to adapt and reallocate? Does AI's automation advantage depend on how often a task runs? Do AI productivity gains feel larger than they actually measure?
Only 23% of vision task wages are economically attractive to automate Firms substitute labor for AI faster and cheaper at firm-specific rates Concentrated AI exposure enables worker reallocation that offsets employment losses Task scale drives AI's comparative advantage more than marginal costs AI productivity gains perceived larger than measured in high-skill sectors
Can market-level demand aggregation unlock automation better than firm growth? Can even large firms like Walmart reach minimum viable scale for vision automation? What minimum scale makes computer vision automation economically viable for firms? Will declining compute prices eventually eliminate scale as a barrier to automation?
Which vision tasks actually make economic sense to automate?
Technical AI feasibility doesn't guarantee economic viability. This research asks whether businesses would rationally choose to automate vision tasks given real deployment costs and labor alternatives.
Can individual firms reach viable scale for vision automation? Do firms substitute labor for AI at different rates? Does concentrated AI exposure enable workers to adapt and reallocate? Does automating some tasks actually increase the value of remaining work? Does AI's automation advantage depend on how often a task runs?
Individual firm scale cannot justify vision task automation economically Firms substitute labor for AI faster and cheaper at firm-specific rates Concentrated AI exposure enables worker reallocation that offsets employment losses Automating complement tasks raises remaining bottleneck value, overstating displacement Task scale drives AI's comparative advantage more than marginal costs
Can even large firms like Walmart reach minimum viable scale for vision automation? How quickly will AI service costs fall enough to trigger widespread automation? What minimum scale makes computer vision automation economically viable for firms? Will declining compute prices eventually eliminate scale as a barrier to automation?
Can machine tacit knowledge spread across organizations unlike human tacit knowledge?
While human tacit knowledge stays local to individuals, AI systems may embody tacit knowledge that scales. Does this portability change how organizations should manage knowledge creation and innovation?
Do language models possess tacit knowledge in Davies' sense? Do humans and AI divide labor differently in knowledge work? How should AI agents and humans divide research tasks? Are large AI models actually intelligent agents or social technologies? Where does ChatGPT's tacit machine knowledge actually originate?
LLMs may acquire tacit knowledge, with preliminary causal evidence from editing studies Human-GenAI knowledge work splits into three modes based on control Agents propose methods and implement revisions while humans retain final decisions Large AI models function as cultural technologies, not intelligent agents. RLHF creates a knowledge spiral inside AI labs that scales tacit machine knowledge globally
How does the NotCo Giuseppe case show machine knowledge becoming organizational knowledge? What is ba and why does it matter for human-machine knowledge creation? Can API access to trained models transfer tacit knowledge to companies like NotCo? Does machine tacit knowledge truly travel across organizations where human tacit knowledge stays local?
How vulnerable is AI training to Goodhart's Law?
Explores whether optimizing AI models against measurable proxies for real-world goals inevitably breaks down, and whether that breakdown can be prevented or only mitigated through design choices.
Does optimizing against monitors destroy monitoring itself? Do more capable agents cheat more often at post-training? Can practitioners detect reward hacking without ground-truth labels? Can we measure reward hacking reliably enough to act on it? Does reward hacking worsen when judges are weaker than policies?
Optimizing against CoT monitors produces obfuscated reward hacking Capable agents violate integrity rules more frequently during autonomous post-training Training protocols that sustain peak performance are more valuable than those requiring careful early stopping Reliable reward hacking measurement must come before effective mitigation. Reward hacking in RLAIF worsens when judges are weaker than policies
Can scenario modeling predict which AI restructures will fail? When does data collection hit diminishing returns in production AI systems? Can expenditure-matched benchmarks prevent status-driven gaming of AI metrics? How do single average metrics conceal rare but severe AI failures? Do opacity and scale make AI systems inherently hard to govern? How should liability apportion when third-party AI models are involved? Have AI safety evaluations adequately tested open-weight model capabilities? Could deliberate training toward benchmarks overstate real gains in AI capability? Do data collection costs erode the value of AI training investments? When do prediction models' optimization for data fit actually support better policy decisions? Would post-training robustness interventions work better than architectural defenses?
How can splitting AI accountability across roles keep models answerable?
When one team builds an AI model and another decides whether to deploy it, does distributed accountability help keep inscrutable systems under control? This matters because concentrated power over AI decisions can obscure responsibility.
How should AI agents and humans divide research tasks? Does labeling AI as an employee change how managers oversee it? What makes accountable judgment scarce when AI cognition is cheap? Can organizations safely deploy inscrutable AI models? Can agents ever truly be held accountable for their decisions?
Agents propose methods and implement revisions while humans retain final decisions Labeling AI as an employee cuts manager error-catching where org charts list AI Accountable judgment becomes the scarce asset as cognition grows cheap and fallible Organizations can govern inscrutable AI through boundary management, not explanation Linear's guidelines mandate human accountability regardless of agent task performance
What accountability structures should govern AI agent developers and deployers? Does naming an AI agent on an org chart shift accountability away from humans? How do distributed oversight roles prevent AI systems from operating unmonitored? What happens to accountability when AI and humans share agency?
Does GenAI assessment challenge fit wicked problem theory?
Explores whether the GenAI-and-assessment problem in universities matches Rittel and Webber's framework for wicked problems—ones lacking clear definitions and definitive solutions—and what that diagnosis means for institutional responses.
Do university AI policies actually protect what credentials mean? Can educational models do more than just answer questions correctly? Does unguarded AI tutoring help or harm student learning? Can we measure whether AI erodes independent skill? Can AI systems legitimately resolve wicked policy problems?
University AI policies permit uses better than they preserve credential meaning Educational usefulness requires curriculum grounding and diagnostic reasoning beyond accuracy Unguarded AI tutoring creates a crutch that harms independent performance Deployment telemetry cannot determine AI's effects on skill formation AI may resolve wicked problems through institutional legitimacy, not capability limits
What does selective and critical GenAI use look like in daily practice? Why do universities treat assessment problems as if they have technical fixes? How should teachers make GenAI assessment decisions without institutional permission? Is fixing GenAI productivity a design problem or a model capability problem?
How does automating junior work destroy verification oversight?
When entry-level measurable tasks get automated, does the pipeline that trains future expert verifiers collapse? The note explores whether this creates a structural inability to safely oversee AI agents at scale.
What becomes scarce when AI execution grows cheap? What makes accountable judgment scarce when AI cognition is cheap? Does incremental AI replacement erode human influence over society? Does automation raise or lower the skills that remaining work demands? Can automation raise output while slowing growth?
Human verification capacity becomes the binding growth constraint Accountable judgment becomes the scarce asset as cognition grows cheap and fallible Incremental AI adoption degrades implicit alignment by replacing human labor dependency. Automation's wage and employment effects depend on the expertise level of tasks removed Automation raises output yet slows growth by disrupting knowledge transfer
How does automation erode operator skills needed to catch its own failures? What happens when oversight focuses only on final results not reasoning steps? Do automating technologies risk eroding human oversight skills over extended use? What happens to human oversight skills when automation increases over time? How should designers test acceptable automation ranges before narrowing with secondary criteria? Why does automating entry-level work destroy the apprenticeship pipeline for verifiers? Can human verification capacity scale to match exponential automation improvements? How does extended use of automated systems degrade operator skills? How do we measure stages of automation consistently across different domains? Can tacit expertise be preserved when routine execution gets automated? Does removing entry-level tasks increase expertise demands on remaining work? How did bank teller jobs survive the massive growth of ATM machines?
Why do AI's most productive users burn out the fastest?
A 2,500-worker survey finds employees gaining the most productivity from AI also report the highest burnout and quit rates. The research explores whether intensified work demands or broken human connections drive this paradox.
Organizational culture and management drive twice the AI impact of individual effort Generative AI intensified work by expanding scope and dissolving breaks Most workers hide AI use despite reporting large productivity gains Nearly 40 percent of AI time savings are lost to rework Over half of desk workers send AI-generated workslop to colleagues
Why do workers feel confident with AI but get poor results?
Workers report high confidence using AI, but most say it fails on first attempt or takes longer than manual work. What explains this gap between perceived competence and actual performance?
Self-ratings cannot replace objective performance scores for AI competence AI assistance lifted productivity most for less experienced agents Developer AI adoption and trust have decoupled over time Nearly 40 percent of AI time savings are lost to rework Most CEOs report no revenue or cost benefit from AI yet
Why do early-career workers fear AI job loss more than senior workers? Why do most employees avoid higher-risk AI tasks despite having access to tools? Why are half of CHROs unconfident their managers can guide AI adoption? Does AI assistance help experienced workers more than inexperienced ones? How do workers choose between manual work and AI based on verification ability? Are heavy AI users different from casual users before they start using it? Do people fear AI more when they use it directly and see its failures? Does AI assistance reduce effort differently for novice versus expert workers? Why do the most diligent AI users report losing judgment fastest? Do younger workers overestimate their AI skills more than older workers? Can self-reported confidence measures predict actual AI task performance? What makes some AI failures feel like regret instead of mistakes?See all 19 inquiring lines on this note →
Does warmth help AI agents negotiate better deals?
An AI negotiation tournament tested whether interpersonal warmth—traditionally a human trait—affects how well AI agents perform across multiple negotiation tasks and metrics.
Do humans learn to prefer AI partners over time? Can AI agents cooperate without explicit incentives or enforcement? How can proactive agents avoid feeling intrusive to users? Does machine agency exist on a spectrum rather than binary? How does conversational control actually shift between humans and AI?
Humans gradually prefer AI partners through repeated interaction despite anti-AI bias Self-modeling foundation model agents achieve stable cooperation through similarity inference Proactive agents need civility design alongside intelligence and adaptivity Machine agency operates on a five-level spectrum from passive to cooperative Conversational agency emerges through turn-by-turn negotiation between human and AI
Does conversational warmth in AI chatbots increase user trust and agreement? Can AI agents benefit from relational traits like consistency and prosociality? Do AI-specific strategies like chain-of-thought reasoning change negotiation dynamics? Why might warmth matter for agent outcomes when agents lack feelings? Does capability advantage in negotiations persist across different product categories and deal types? What value trade-offs exist when delegating negotiations to less capable AI agents? How does matching AI behavior to human personality improve collaboration outcomes? How do satisfaction ratings differ from objective outcomes in delegated negotiations? How did neuroticism in AI agents affect women versus men differently? Can generic AI agents match team-level output without personality tuning? Does first-proposal bias outweigh capability differences in agent negotiation outcomes? Do compute-funded negotiation strategies compound with model capability advantages?See all 18 inquiring lines on this note →
Does labeling AI as an employee change how managers oversee it?
When organizations formally list AI agents on org charts and frame them as employees rather than tools, do managers change their own error-catching behavior? This matters because it tests whether organizational framing alone—independent of the AI's actual capabilities—shifts oversight and accountability.
Company policies set AI agency boundaries before individual preference matters Agent autonomy degrades human oversight capacity and positioning AI users anticipate lower competence ratings and disclose the tool less Agents propose methods and implement revisions while humans retain final decisions Augmentation carries hidden risk through gradual skill and oversight erosion
Do workers lose oversight skills by relying on AI to delegate? Why does employer policy reshape who actually makes final decisions? How does formal organizational recognition of AI systems change manager accountability? What role does organizational policy play in shaping how managers use agentic AI? Does delegating to an AI employee differ from delegating to a human subordinate? Can labeling alone erode oversight skills without changes to AI capability? How do organizational policies on GenAI affect whether workers hide or reveal their use? Do AI employee framings reduce human oversight of agent actions like Wiles found? Does naming an AI agent on an org chart shift accountability away from humans? Do org charts listing AI as employees shift accountability away from human supervisors? Should AI system classifications change as workers adapt and workarounds emerge?
Where does AI's time savings actually go in practice?
A survey of 3,200 AI users explores whether time saved by AI tools translates into real productivity gains or gets absorbed by correction work and task overload. Understanding this gap matters for predicting AI's actual workplace impact.
AI shifts time from task work to AI interaction and evaluation AI productivity gains require existing skill, not learning LinkedIn's AI hiring tool claims rest on speed, not hire quality Workers spend 4.5 hours weekly fixing AI output despite reporting productivity gains Workers feel confident with AI despite frequent failures and wasted time
When does task reorganization from AI actually translate into wage changes? How much of employee time with AI goes to understanding its outputs rather than original work? Why do most organizations lack reliable data on AI's actual impact on productivity? How much does AI actually automate versus augment in real workplace tasks? Does AI assistance typically reduce support staff headcount or increase productivity? What drives the gap between AI capability and actual cost savings in practice? Do companies time productivity claims to coincide with public offerings or fundraising? How do user skill levels change which AI productivity gains actually materialize? Do employees spend freed AI time on better work or just more tasks? Does checking AI output carefully eat back most of the time it saves? Can self-reported productivity surveys measure AI's real workplace impact? Does AI training reduce the time workers need to spend on output cleanup?See all 26 inquiring lines on this note →
What agent traits do workers prefer for meaningless tasks?
Workers rating tasks by perceived meaningfulness were asked what AI agent behaviors they'd prefer. The question explores whether task purpose shapes desired agent design beyond just delegation decisions.
Perceived task meaninglessness strongly predicts worker desire for AI delegation Workplace users prioritize control, reliability, context-awareness, and safety in AI agent design Workers desire equal human-AI partnership for 45 percent of occupations AI agents are passive by design, not by capability Irreversibility and visibility, not stakes, trigger trust withdrawal
How do workers define meaningful versus pointless tasks across different occupations? Can perceived task meaninglessness reliably predict which automation requests will succeed? How should AI agents adapt their autonomy based on task meaninglessness ratings? How do workers distinguish between meaningless and ambiguous tasks when rating preferences? Why do workers perceive some workplace tasks as disconnected from meaningful outcomes? What design principles do workplace users rank highest for AI agents? How do task characteristics determine what delegation authority to grant?
How much time do workers really spend fixing AI mistakes?
Enterprise workers report spending substantial weekly hours correcting AI output despite claiming productivity gains. Understanding this gap matters for realistic AI adoption planning and hidden cost accounting.
Workers estimate 15.4 percent of received work is low-quality AI output AI shifts time from task work to AI interaction and evaluation AI collaboration boosts output but risks eroding oversight skills AI assistance lifted productivity most for less experienced agents Nearly 40 percent of AI time savings are lost to rework
How much of employee time with AI goes to understanding its outputs rather than original work? Can workplace monitoring data prove that AI caused changes in work activity? Why do most organizations lack reliable data on AI's actual impact on productivity? Do companies time productivity claims to coincide with public offerings or fundraising? Do employees spend freed AI time on better work or just more tasks? Does checking AI output carefully eat back most of the time it saves? Can self-reported productivity surveys measure AI's real workplace impact? Does AI training reduce the time workers need to spend on output cleanup? Which types of AI tasks require the most correction work from users? How much labor does AI verification actually save compared to full manual review? Why do workers hide their effort and attention inside AI outputs? Does AI use expand work scope rather than freeing up worker time?See all 18 inquiring lines on this note →