Can AI learn social norms better than humans?
Explores whether large language models can predict cultural appropriateness more accurately than individual humans, and what this reveals about how social knowledge is transmitted and learned.
Hook: GPT-4.5 is better at knowing what's socially appropriate than any individual human. Not some humans — all of them. 100th percentile. But it makes mistakes that every other AI model also makes in the same way.
The finding:
555 everyday scenarios. "How appropriate is it to laugh at a job interview?" "To cry on a bus?" "To read in church?" When asked to predict the average human judgment, GPT-4.5 was more accurate than every single human participant. Replicated with Gemini 2.5 Pro (98.7%), GPT-5 (97.8%), Claude Sonnet 4 (96.0%).
The AI doesn't just know the rules. It knows the collective sense of a culture better than the people living in it.
Why this matters:
The dominant theory in cognitive science says social norms require embodied experience — you learn what's appropriate by living in a culture, reading faces, feeling social consequences. Statistical learning over text shouldn't be enough. But it is. "Sophisticated models of social cognition can emerge from statistical learning over linguistic data alone."
Language turns out to be a "remarkably rich repository for cultural knowledge transmission." Everything humans write — from etiquette guides to Reddit arguments to novels — encodes social norms. The AI has read more of this than any human could experience in a lifetime.
The catch:
All models show "systematic, correlated errors." Not random mistakes — structured blind spots that every AI architecture shares. The same scenarios that trip up GPT-4.5 also trip up Gemini and Claude. This pattern "indicates potential boundaries of pattern-based social understanding."
There are aspects of social norms that don't make it into text. The unwritten rules that communities enforce through glances, silences, and physical presence. The norms that are so obvious nobody bothers to articulate them. These are the correlated blind spots — and they're exactly the norms you most need to get right in practice.
The tension:
The AI is a savant — extraordinary competence in one dimension (predicting collective norms from text) combined with systematic gaps in another (the norms that never get written down). Better than any individual at the average, blind to the specifics that any local participant would catch immediately.
Flat, not targeted — the post-generation consequence. The savant-from-outside pattern has a specific consequence at the level of generated posts: AI output is flat rather than targeted because no social position is occupied. Normal influencer, commentator, and pundit speech online carries implicit position-taking that situates the speaker relative to the audience — speaking as one of us, or for this community, or against that one. The position-taking is what makes the content addressed to someone in particular, rather than written about a topic in general. AI can predict the average appropriate response but cannot occupy a specific social position vis-à-vis a specific community, because it has no community membership to mark. The output is therefore flat — competent on general norm, absent on the position-taking that would make the post legible as speech from someone to someone. Knowing norms from outside and speaking from outside produce the same residue: content that is addressed to no one in particular and therefore cannot perform the community-specific legitimacy that targeted commentary depends on.
Post structure: Hook (the number) → What it means (embodiment challenge) → The catch (correlated errors) → The tension (savant pattern) → What this means for AI deployment in social contexts
Platform: LinkedIn (300-400 words, practical tone) or Medium (longer with theoretical framing)
Inquiring lines that read this note 80
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.
What should agent evaluation prioritize to reveal reliable behavior? How can reward models capture diverse human preferences without excluding minority populations? How should designers communicate what AI systems truly are and can do?- Why can't AI models internalize audiences the way human experts do?
- What genuine cultural forms does AI homogeneity actually displace?
- What distinguishes genuine cultural understanding from exploited surface-level elimination strategies?
- How do humans and AI develop accurate models of each other?
- Can humans learn accurate models of AI through repeated interaction without labels?
- Can AI systems recognize intelligence in humans the way humans recognize it in each other?
- What role does bidirectional model updating play in human-AI understanding?
- Will AI saturation push discourse toward oral culture's strengths and weaknesses?
- How does communicative standing depend on participation in normative communities?
- Do language models calibrate to actual human pragmatic norms?
- What expectations does human conversation activate that AI should avoid triggering?
- Can AI models predict whether alignment reads as warmth versus mockery in different cultures?
- Do AI models accurately predict what is socially appropriate in human conversations?
- Can AI safely personalize within negotiated societal bounds?
- What social patterns from human training data activate in agent context?
- Why do moderately represented cultures show more flattening than data-poor cultures?
- Can AI predict social norms well enough without embodied experience?
- How does community validation shape unconventional human-AI relationships?
- How do AI errors in norm prediction differ from systematic human errors?
- Does predicting social norms from outside count as participation?
- Can automated systems encode human values as reliably as human workers enforce them?
- How much does demographic bias in guardrails mirror real-world social inequalities?
- Do culturally distinct human groups create similar attribution errors as human-AI mixtures?
- Can individually accurate agents still fail at population-level representation?
- How do language models predict collective social norms better than individual humans?
- How do cultural norms reshape initial interpretations of social intent?
- Do agents develop genuine social behavior despite interaction density?
- How do AI models balance competing social goals simultaneously?
- Do AI systems need embodiment to understand social norms?
- Why do standard social regularization methods miss the actual value networks provide?
- Do different AI models independently converge on the same social outputs?
- What social norms do AI systems consistently fail to understand?
- How much cultural knowledge exists only in unwritten social rules?
- Should simulated value-based discussions be validated against real human populations?
- Do language models understand tacit workplace norms and unspoken social rules?
- Does social grounding in language improve through iterative human integration?
- Can language models develop genuine social grounding through human interaction?
- Can LLMs predict social norms without deep integration into linguistic practices?
- Why do language models respond to human social influence patterns?
- How do low-dimensional representation structures entangle multiple cultures together?
- Can statistical learning from language alone capture all aspects of cultural competence?
- Why do automated selection methods outperform human judgments of relevant context?
- Can statistical learning from text replace embodied cultural experience?
- What social information is missing from language data?
- How should AI systems model human resource constraints and expertise levels?
- Should AI assistants align with role-specific norms rather than user preferences?
- Should AI alignment follow individual preferences or role-based norms?
- Does approaching human performance mean learning the same grammatical rules?
- Can large language models predict social norms better than individual script variation?
- Do language models systematically overestimate accuracy on collective behavior tasks?
- Why do language models approximate collective human judgment better than individuals?
- Do individual language models match particular human judges better than population averages?
- What role does contingent interaction play in activating social response norms?
- Can proactive AI agents deploy politeness strategies without appearing intrusive?
- What social boundaries must proactive agents respect during conversation?
- How does an AI agent's autonomy level interact with its social cues?
- Where does AI's communicative agency fall on spectrums beyond the passive-active binary?
- Should AI alignment use normative standards instead of aggregate preferences?
- Can AI-assisted alignment eventually solve fairness at scale?
- Does alignment training create the shared prosocial pattern across models?
Related concepts in this collection 4
This note in its neighbourhood — explore the map, then jump to a related concept in the list below.
Click a node to walk · click center to open · click Open in graph to see this note in the full knowledge graph
-
Can AI systems learn social norms without embodied experience?
Large language models exceed individual human accuracy at predicting collective social appropriateness judgments. Does this reveal that embodied experience is unnecessary for cultural competence, or do systematic AI failures point to limits of statistical learning?
primary evidence
-
What makes linguistic agency impossible for language models?
From an enactive perspective, does linguistic agency require embodied participation and real stakes that LLMs fundamentally lack? This matters because it challenges whether LLMs can truly engage in language or only generate text.
the theory being challenged
-
Can LLMs acquire social grounding through linguistic integration?
Explores whether LLMs gradually develop social grounding as they become embedded in human language practices, analogous to child language acquisition. Tests whether grounding is a fixed property or an outcome of participatory use.
complicates the trajectory: maybe grounding is already sufficient for norm prediction
-
Can AI agents learn people better from interviews than surveys?
Can rich interview transcripts seed more accurate generative agents than demographic data or survey responses? This matters because it challenges how we build digital simulations of real people.
complementary evidence for the post angle: social norm prediction at 100th percentile + interview-based response replication at 85% demonstrate text-based learning approximates embodied social knowledge across different task dimensions
Related papers in this collection 8
Papers most semantically related to this note, ranked by cosine similarity in the embedding space.
- AI Models Exceed Individual Human Accuracy in Predicting Everyday Social Norms
- Humans learn to prefer trustworthy AI over human partners
- SOTOPIA: Interactive Evaluation for Social Intelligence in Language Agents
- SDPO: Segment-Level Direct Preference Optimization for Social Agents
- On Epistemic Diversity in Large Language Models
- Incoherent by Design? On the Moral Self-Consistency of LLMs
- MetaMind: Modeling Human Social Thoughts with Metacognitive Multi-Agent Systems
- The Goldilocks of Pragmatic Understanding: Fine-Tuning Strategy Matters for Implicature Resolution by LLMs
Original note title
the social norm savant — ai knows your culture better than you do but from the outside