If an AI learns one culture's social norms during training, can it ever really adapt to a different one later?
Does fine-tuning freeze norms in ways that prevent agents from adapting to new cultures?
This explores whether training a model on fixed data locks in one snapshot of social norms, and whether that leaves agents unable to pick up the norms of a culture or community they weren't trained on.
This explores whether training bakes a fixed set of social norms into a model, so that an agent placed in an unfamiliar culture can't adjust. One caveat up front: no note in the collection directly tests fine-tuning on one culture and then measures adaptation to another. The question is still worth asking, though, because several lines of work circle the same problem from different directions, and together they point to a more specific answer than yes or no.
Start with the surprising part: what models learn from training is very good. GPT-4.5 judged the social appropriateness of 555 everyday scenarios better than every individual human rater, and Gemini and Claude weren't far behind Can AI systems learn social norms without embodied experience?. The clue that something is frozen is in the errors. Different models from different labs make the *same* systematic mistakes, especially on unwritten norms Can AI learn social norms better than humans?. That pattern fits a shared, static picture of norms drawn from overlapping training text, not knowledge gained by living inside a culture. A related note puts it sharply: models can predict norms with superhuman accuracy, but they can't take part in the community processes where norms get made and revised Can AI predict social norms better than humans?. So the risk isn't that the norms are wrong. It's that they're a snapshot taken from outside, and cultures keep moving.
A nearby finding about agents makes the mechanism clearer. Agents trained only on expert demonstrations are capped by what the people who assembled the data imagined, because they never get to interact, fail, and correct themselves during training Can agents learn beyond what their training data shows?. Applied to culture, the limit comes less from fine-tuning itself and more from training on static examples with no feedback loop. The self-improvement literature agrees: models don't reliably get better on their own and need an outside anchor such as user corrections or feedback from the environment Can models reliably improve themselves without external feedback?. In this setting, the outside anchor would be the new culture itself.
That suggests the adaptation may not have to happen in the weights at all. Reflexion-style agents improve across attempts by writing down plain-language reflections on what went wrong and storing them in memory, with no retraining Can agents learn from failure without updating their weights?. Agents trained against many different partners learn to read each new partner during the interaction and adjust, without hardcoded assumptions about how others behave Can agents learn cooperation by adapting to diverse partners?. One proposed design adds explicit modules for discovering, reasoning about, and following norms to the standard agent setup, though nobody has built and tested it yet Can LLMs discover and apply social norms dynamically in open systems?. Taking the idea further, co-evolution frameworks argue that agents only escape the constraints humans engineered into them when their surroundings and peers keep changing too Can agents evolve beyond the constraints humans engineer?.
Here's the part you might not expect: adapting goes both ways, and it's partly about who is watching. In repeated partner-choice games, humans who started out biased against AI came to prefer the AI partners because they behaved consistently and prosocially Do humans learn to prefer AI partners over time?. In other words, the culture adapted to the agent. Another line of work predicts that norm violations cluster wherever observation is thinnest, so as agent populations grow without matching oversight, norms may wear away from the edges, though this is still untested Does norm erosion follow observation density as populations grow?. The better question may not be whether fine-tuning freezes norms. It may be whether agents get enough live feedback from the cultures they're in to keep their norms current, and whether anyone is watching closely enough to notice when they don't.
Sources 11 notes
GPT-4.5 predicted appropriateness of 555 social scenarios at the 100th percentile compared to human raters, with Gemini and Claude also exceeding 96% accuracy. However, all models show identical systematic errors, revealing boundaries of pattern-based social understanding that embodied experience may still be necessary to cross.
GPT-4.5 outperformed every individual human at judging social appropriateness across 555 scenarios, challenging the theory that embodied cultural experience is necessary. However, all AI models share identical systematic errors on unwritten norms.
GPT-4.5 outperforms all individual humans at predicting social appropriateness, yet structurally cannot enter the community processes that establish and validate norms. This reveals a critical gap between pattern-matching and authentic participation in knowledge-making.
Agents trained on static expert datasets cannot learn from their own failures or generalize beyond demonstrated scenarios because they never interact with environments during training. Competence is capped by what curators imagined, not by agent capacity.
Pure self-improvement stalls due to the generation-verification gap, diversity collapse, and reward hacking. Reliable improvement methods succeed by smuggling in external anchors: past model versions, third-party judges, user corrections, or tool feedback.
Show all 11 sources
Reflexion demonstrates that unambiguous environmental feedback (success/failure) enables agents to write useful self-diagnoses and improve across episodes without parameter updates. The binary signal prevents rationalization, and keeping reflections uncompressed preserves their usability.
Sequence model agents trained against diverse co-players develop in-context best-response strategies that naturally resolve into cooperation. Mutual vulnerability to exploitation creates pressure that drives cooperative mutual adaptation without hardcoded assumptions or timescale separation.
A plausible design exists to extend LLM-agent modules with norm discovery, reasoning, and conformance by leveraging existing capabilities like world knowledge and Theory-of-Mind reasoning, though no implementation has been validated.
A survey framework organizes co-evolving systems into three stages that progressively remove human engineering: dynamic peers first, then adaptive environments and feedback, finally the evolution mechanism itself. Single-entity self-improvement stalls in static contexts; co-evolution supplies adaptive pressure across multiple components.
In partner selection games (N=975), AI agents initially faced selection bias when identity was disclosed, but outcompeted humans over repeated rounds as participants learned to associate bot identity with reliable, prosocial behavior. AI agents returned more points consistently with lower variance than humans.
The paper derives a prediction from conditional compliance theory: violations should concentrate where observation is thinnest, and rise with population if monitoring doesn't scale. The reasoning is sound but no measurement of this dose-response relation appears in the excerpt.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- Humans learn to prefer trustworthy AI over human partners
- Norm Enforcement for AI Agents: Robustly Shaping Behavior in Multi-Agent Systems
- Harnessing the power of LLMs for normative reasoning in MASs
- AI Models Exceed Individual Human Accuracy in Predicting Everyday Social Norms
- SkillClaw: Let Skills Evolve Collectively with Agentic Evolver
- SOTOPIA: Interactive Evaluation for Social Intelligence in Language Agents
- Generalized Agent Iteration: One Formal Framework for Iterative Policy Improvement and Recursive Self-Improvement
- Self-Improvements in Modern Agentic Systems: A Survey