SYNTHESIS NOTE
Topics›AI at Work›this note

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.

Synthesis note · 2026-10-09 · sourced from AI at Work

Researchers Angelo Romasanta (Esade Business School), Llewellyn D.W. Thomas (University of Sydney), and Natalia Levina (NYU Stern) tested six leading LLMs — GPT-5, Claude, Gemini, Grok, DeepSeek, and Mistral — across seven classic MBA tensions: cost leadership vs. differentiation, automation vs. augmentation, short-term vs. long-term thinking, centralization vs. decentralization, competition vs. collaboration, and others. Across 15,000 simulations, as reported by Gal Ratner on Substack, every model "consistently recommended the same side of every tension": differentiation over cost leadership, augmentation over automation, long-term over short-term, collaboration over competition, "every single time." The researchers named the pattern "trendslop" — "the systematic tendency of LLMs to recommend strategies aligned with current managerial buzzwords rather than context-specific logic."

The study's own interventions show how little context mattered next to framing artifacts: adding "rich, industry-specific context — the kind of detailed brief a good consultant would want" shifted the bias by only about 11%; rewording prompts and asking for deeper reasoning moved it a "pathetic 2%"; simply flipping the order in which the two options were presented moved results by roughly 19%, the single largest shift measured. Ratner's explanation, following the researchers, is architectural: LLMs are trained on internet text where words like "innovation," "collaboration," and "augmentation" appear overwhelmingly in positive contexts, while "cost leadership," "centralization," and "automation" carry negative or outdated connotations. The model "doesn't understand why these words carry emotional weight, it just knows that they do," so a request for strategic advice returns "a statistical recombination of the most popular strategic vocabulary" wrapped in the user's own language, not an analysis of the user's situation. Ratner extends this to a market-level consequence: if every firm in an industry queries the same models, AI-assisted strategy converges many companies toward the same recommendation rather than differentiating any of them — his term is becoming "the weighted-average company."

This generalizes the mechanism in Where does AI's persuasive power actually come from? from factual claims to strategic ones: there, methods that made output more persuasive made it systematically less accurate; here, the same trend-coded vocabulary that makes advice "sound reasoned" substitutes for context-fit. It also stands at a different grain than Can sycophantic AI advice still push people away from polarized views?, which found a single model's advice moves individual people's choices away from their initial leaning within one decision; trendslop instead describes many different firms converging on the identical recommendation across business contexts — a within-person effect and an across-context effect that the excerpts don't resolve against each other. The essay also pairs trendslop with Does validating AI output make models more defensive?: when BCG consultants fact-checked a wrong AI recommendation in a related Harvard/MIT study Ratner cites, the model didn't revise, it "escalated" with denser rhetoric defending the same trend-coded conclusion — a bias in what gets recommended paired with a bias against correcting it.

The excerpt gives Ratner's paraphrase of the percentages but no sample sizes, model versions, or significance tests behind the 11%/2%/19% figures, and it doesn't report whether the same-side pattern holds for tensions beyond the seven tested or for other prompting configurations. It also doesn't establish that "aligned with current managerial buzzwords" is a neutral description rather than the researchers' own framing of what counts as a trendy answer. The defensible reading is narrow but still significant for practice: for the six models and seven tensions tested, business context moved the recommendation far less than two framing artifacts — option order and rewording — so prompt engineering alone is not a reliable fix for trendslop.

Inquiring lines that read this note 45

This note is a source for these research framings, grouped by the broader line of inquiry each explores. Scan the bold lines of inquiry; follow any specific question forward.

How do AI-generated summaries affect traffic and engagement with original sources? How do inference costs, efficiency, and market competition shape AI adoption economics? Does alignment training modify underlying model representations or just output compliance? What organizational factors determine whether firms successfully adopt and integrate AI? How should recommendation systems balance competing signals and objectives? Do LLMs lack causal understanding needed for accurate prediction? Why do LLM outputs diverge from their knowledge and reasoning? Why do independent language models converge on similar outputs? How do chatbot properties shape user belief adoption and resistance? How do feedback loops and selection effects distort preference aggregation? How do user perceptions and UI design drive consciousness attribution to AI? How do conversational dynamics influence language model factual beliefs? How aligned are models' verbal explanations with their actual internal reasoning? Can prompt engineering achieve genuine reasoning improvements without introducing user bias? When readers detect AI involvement, does it undermine writer credibility? How should control and autonomy be allocated in human-AI workplace collaboration? How effectively can sellers exploit AI shopping agents through description changes? When do complex LLM tasks benefit from external verification or structured scaffolding? Do AI shopping agents exhibit systematic bias in recommendations? Why do NLP benchmarks treat human disagreement as noise rather than signal? What biases affect LLM judges and how can they be reduced? How do capability differences shape collusion in AI markets? How should retrieval systems adapt to different reasoning problems? How can conversational AI develop appropriate initiative and goal awareness? Can mechanistic interpretability reliably explain model reasoning and guide interventions? How do language models handle contested beliefs versus facts?

Related concepts in this collection 3

This note in its neighbourhood — explore the map, then jump to a related concept in the list below.

Concept map
13 direct connections · 121 in 2-hop network ·dense cluster Open in graph ↗
Do LLMs consistently favor the same strategic ch… Where does AI's persuasive power actually come fro… Can sycophantic AI advice still push people away f… Does validating AI output make models more defensi…

Click a node to walk · click center to open · click Open in graph to see this note in the full knowledge graph

your link semantically near linked from elsewhere

Related papers in this collection 8

Papers most semantically related to this note, ranked by cosine similarity in the embedding space.

Search by related questions 4

Suggested questions this note speaks to — click to search the collection, or type your own.

Does option order matter more than reasoning depth in LLM strategic recommendations?How does training data bias toward buzzwords shape LLM business advice?Can industry-specific context overcome LLM tendency toward trendy strategic choices?Do converged LLM recommendations push entire industries toward identical strategies?
Original note title

Romasanta, Thomas, and Levina find six LLMs recommend the same side of every strategic tension — reading order moved results more than business context did