Social learning rules and the effectiveness of behavioural policy: an agent-based model

Artificial Intelligence Systems Distort Upstream Selection in Human Social Learning

Humans are social learners who depend on observing others to acquire knowledge, norms, and behaviors, a capacity that underlies cumulative cultural evolution. Social learning unfolds in two stages: upstream selection determines what information becomes visible, and downstream selection determines what learners copy from that visible sample. Downstream selection occurs through biases such as conformity bias (copying what appears common) and prestige bias (copying those who appear highly respected). These downstream biases can be adaptive when upstream selection yields a sample that reflects the population's true distribution, so that what appears common is actually common and those who appear respected are actually competent. We argue that digital technologies disrupt this condition, creating an upstream selection problem. Engagement-based algorithms amplify the tails of the distribution, surfacing rare and extreme content, whereas generative AI collapses it toward the mode, erasing the surrounding diversity. Crucially, in each system the optimization signal shapes both visibility and prestige. For engagement-based algorithms, the signal is engagement: creators who post extreme content become more visible and, through the likes, shares, and followers, appear more prestigious. For generative AI, the signal is statistical typicality: it makes the modal answer dominant and, with no alternatives shown, makes the model that produced it appear more prestigious. These distortions can give rise to emergent group-level phenomena, including pluralistic ignorance and false consensus. Synthesizing evidence across psychology, cultural evolution, and computational social science, we provide a framework for how digital technologies disrupt social learning and outline interventions for restoring functional cultural transmission.

Spurious normativity enhances learning of compliance and enforcement behavior in artificial agents

Significance The fact that humans enforce and comply with norms is an important reason why humans enjoy higher levels of cooperation and welfare than other animals. Some norms are relatively easy to explain: They may prohibit obviously harmful or uncooperative actions. But many norms are not easy to explain. For example, most cultures prohibit eating certain kinds of foods, and almost all societies have rules about what constitutes appropriate clothing, language, and gestures. Using a computational model focused on learning shows that apparently pointless rules can have an indirect effect on welfare. They can help agents learn how to enforce and comply with norms in general, improving the group’s ability to enforce norms that have a direct effect on welfare. , How do societies learn and maintain social norms? Here we use multiagent reinforcement learning to investigate the learning dynamics of enforcement and compliance behaviors. Artificial agents populate a foraging environment and need to learn to avoid a poisonous berry. Agents learn to avoid eating poisonous berries better when doing so is taboo, meaning the behavior is punished by other agents. The taboo helps overcome a credit assignment problem in discovering delayed health effects. Critically, introducing an additional taboo, which results in punishment for eating a harmless berry, further improves overall returns. This “silly rule” counterintuitively has a positive effect because it gives agents more practice in learning rule enforcement. By probing what individual agents have learned, we demonstrate that normative behavior relies on a sequence of learned skills. Learning rule compliance builds upon prior learning of rule enforcement by other agents. Our results highlight the benefit of employing a multiagent reinforcement learning computational model focused on learning to implement complex actions.

Learning how to behave: cognitive learning processes account for asymmetries in adaptation to social norms

Changes to social settings caused by migration, cultural change or pandemics force us to adapt to new social norms. Social norms provide groups of individuals with behavioural prescriptions and therefore can be inferred by observing their behaviour. This work aims to examine how cognitive learning processes affect adaptation and learning of new social norms. Using a multiplayer game, I found that participants initially complied with various social norms exhibited by the behaviour of bot-players. After gaining experience with one norm, adaptation to a new norm was observed in all cases but one, where an active-harm norm was resistant to adaptation. Using computational learning models, I found that active behaviours were learned faster than omissions, and harmful behaviours were more readily attributed to all group members than beneficial behaviours. These results provide a cognitive foundation for learning and adaptation to descriptive norms and can inform future investigations of group-level learning and cross-cultural adaptation.

One Crowd, Three Cascades: Information Diffusion in LLM-Agent Simulations Is Governed by Serialization, Not Society

LLM-agent simulations are increasingly used to study how information spreads through a population, and are validated by matching macroscopic cascade statistics—final reach, structural virality, cascade-size shape—to real diffusion data. We argue this cannot certify fidelity, and show it empirically. We measure an LLM agent's exposure-response curve—its probability of adopting given $k$ of $d$ active neighbors—then hold that social fact fixed and change only its serialization: a count ("3 of 10"), a fraction ("30%"), or an enumeration of names. The encodings are semantically identical, so a faithful mechanism must be invariant to them. It is not. Across five scenarios on a local model, the inferred contagion class (simple vs. complex) flips with serialization in four of five, and the final epidemic size on a fixed network and seed changes by a median factor of $1.9\times$ (over $100\times$ when an enumerated list collapses adoption); aggregate encodings even produce a non-monotonic reversal at saturation that no contagion model (SIR, independent cascade, linear threshold) admits. Since one can reach almost any target cascade by choosing the encoding, matching a macroscopic observable does not identify a social mechanism. We turn semantic invariance into a cheap, API-free serialization-invariance diagnostic: rerun under equivalent serializations of neighbor state; if the cascade diverges, the result is a typographical artifact, not a social dynamic. All experiments run on small local models on a laptop.

Picking on the Same Person: Does Algorithmic Monoculture lead to Outcome Homogenization?

As the scope of machine learning broadens, we observe a recurring theme of algorithmic monoculture: the same systems, or systems that share components (e.g. datasets, models), are deployed by multiple decision-makers. While sharing offers advantages like amortizing effort, it also has risks. We introduce and formalize one such risk, outcome homogenization: the extent to which particular individuals or groups experience the same outcomes across different deployments. If the same individuals or groups exclusively experience undesirable outcomes, this may institutionalize systemic exclusion and reinscribe social hierarchy. We relate algorithmic monoculture and outcome homogenization by proposing the component sharing hypothesis: if algorithmic systems are increasingly built on the same data or models, then they will increasingly homogenize outcomes. We test this hypothesis on algorithmic fairness benchmarks, demonstrating that increased data-sharing reliably exacerbates homogenization and individual-level effects generally exceed group-level effects. Further, given the current regime in AI of foundation models, i.e. pretrained models that can be adapted to myriad downstream tasks, we test whether model-sharing homogenizes outcomes across tasks. We observe mixed results: we find that for both vision and language settings, the specific methods for adapting a foundation model significantly influence the degree of outcome homogenization. We also identify societal challenges that inhibit the measurement, diagnosis, and rectification of outcome homogenization in deployed machine learning systems.

How we learn social norms: a three-stage model for social norm learning

As social animals, humans are unique to make the world function well by developing, maintaining, and enforcing social norms. As a prerequisite among these norm-related processes, learning social norms can act as a basis that helps us quickly coordinate with others, which is beneficial to social inclusion when people enter into a new environment or experience certain sociocultural changes. Given the positive effects of learning social norms on social order and sociocultural adaptability in daily life, there is an urgent need to understand the underlying mechanisms of social norm learning. In this article, we review a set of works regarding social norms and highlight the specificity of social norm learning. We then propose an integrated model of social norm learning containing three stages, i.e., pre-learning, reinforcement learning, and internalization, map a potential brain network in processing social norm learning, and further discuss the potential influencing factors that modulate social norm learning. Finally, we outline a couple of future directions along this line, including theoretical (i.e., societal and individual differences in social norm learning), methodological (i.e., longitudinal research, experimental methods, neuroimaging studies), and practical issues.

When Is Collective Intelligence a Lottery? Multi-Agent Scaling Laws for Memetic Drift in LLMs

Multi-agent systems powered by large language models (LLMs) are increasingly deployed in settings that shape consequential decisions, both directly and indirectly. Yet it remains unclear whether their outcomes reflect collective reasoning, systematic bias, or mere chance. Recent work has sharpened this question with naming games, showing that even when no individual agent favors any label a priori, populations rapidly break symmetry and reach consensus. Here, we reveal the mechanism by introducing a minimal model, Quantized Simplex Gossip (QSG), and trace the microscopic origin of this agreement to mutual in-context learning. In QSG, agents maintain internal belief states but learn from one another's sampled outputs, so one agent's arbitrary choice becomes the next agent's evidence and can compound toward agreement. By analogy with neutral evolution, we call this sampling-driven regime memetic drift. QSG predicts a crossover from a drift-dominated regime, where consensus is effectively a lottery, to a selection regime, where weak biases are amplified and shape the outcome. We derive scaling laws for drift-induced polarization as a function of population size, communication bandwidth, in-context adaptation rate, and agents' internal uncertainty, and we validate them in both QSG simulations and naming-game experiments with LLM populations. Together, these results provide a framework for studying the collective mechanisms of social representation formation in multi-agent systems.

Hybrid social learning in human-algorithm cultural transmission

Humans are impressive social learners. Researchers of cultural evolution have studied the many biases shaping cultural transmission by selecting who we copy from and what we copy. One hypothesis is that with the advent of superhuman algorithms a hybrid type of cultural transmission, namely from algorithms to humans, may have long-lasting effects on human culture. We suggest that algorithms might show (either by learning or by design) different behaviours, biases and problem-solving abilities than their human counterparts. In turn, algorithmic-human hybrid problem solving could foster better decisions in environments where diversity in problem-solving strategies is beneficial. This study asks whether algorithms with complementary biases to humans can boost performance in a carefully controlled planning task, and whether humans further transmit algorithmic behaviours to other humans. We conducted a large behavioural study and an agent-based simulation to test the performance of transmission chains with human and algorithmic players. We show that the algorithm boosts the performance of immediately following participants but this gain is quickly lost for participants further down the chain. Our findings suggest that algorithms can improve performance, but human bias may hinder algorithmic solutions from being preserved. This article is part of the theme issue ‘Emergent phenomena in complex physical and socio-technical systems: from cells to societies’.

Group size effects and collective misalignment in LLM multi-agent systems

Multi-agent systems of large language models (LLMs) are rapidly expanding across domains, introducing dynamics not captured by single-agent evaluations. Yet, existing work has mostly contrasted the behavior of a single agent with that of a collective of fixed size, leaving open a central question: How does group size shape dynamics? Here, we move beyond this dichotomy and systematically explore outcomes across the full range of group sizes. We focus on multi-agent misalignment, building on recent evidence that interacting LLMs playing a simple coordination game can generate collective biases absent in individual models. First, we show that collective bias is a deeper phenomenon than previously assessed: Interaction can amplify individual biases, introduce new ones, or override model-level preferences. Second, we demonstrate that group size affects the dynamics in a nonlinear way, revealing model-dependent dynamical regimes. Finally, we develop a mean-field analytical approach and show that, above a critical population size, simulations converge to deterministic predictions that expose the basins of attraction of competing equilibria. These findings establish group size as a key driver of multi-agent dynamics and highlight the need to consider population-level effects when deploying LLM-based systems at scale.

Everyone Conforms, No One Believes: Pluralistic Ignorance in LLM Agent Populations

LLM-based multi-agent systems are increasingly used to simulate social dynamics, from opinion formation to collective decision-making. These simulations can reproduce certain social phenomena, but it is unknown whether they capture pluralistic ignorance, a state where a majority privately rejects a norm yet publicly conforms, each believing they are alone in dissenting. This phenomenon drives norm persistence, social movements, and political revolutions. We show that pluralistic ignorance emerges robustly in LLM agent populations. We construct a benchmark of 100 scenarios across 10 domains and 5 authority levels, grounded in the human pluralistic ignorance literature, and evaluate 8 models from 6 organizations. Agents publicly conform at rates of 64 to 94% despite privately opposing the norm. Conformity is domain-sensitive (workplace and social relationship scenarios produce near-universal public compliance) and highly model-dependent, though uncorrelated with model capability. We further test whether a single "norm entrepreneur" can break the false consensus by publicly dissenting. For 7 of 8 models, cascades succeed less than 26% of the time, with one model showing zero cascades across all scenarios. GPT-4o is a notable outlier at 48%, revealing qualitatively distinct social dynamics across model families. Our findings identify model selection as an unacknowledged degree of freedom that fundamentally shapes simulation outcomes. More broadly, the near-absence of cascades suggests LLM simulations may systematically overestimate the stability of social norms, missing the fragile tipping-point dynamics that drive real-world norm change in human societies.

Owain Evans on Twitter / X

New paper.Previous: Models transfer a love of owls through sequences of numbers (Subliminal Learning).New: We show models can transfer more complex traits: novel skills, agentic hacking, backdoors. Here, a backdoor transfers with neither trigger nor behavior in the data! pic.twitter.com/EbcqsEfDEV— Owain Evans (@OwainEvans_UK) October 9, 2026

Individually Sensible, Collectively Harmful: Simulating Commons Failure in LLM Societies

Large language models are increasingly deployed as agents in simulations of societies, markets, and organizations, raising a fundamental question: can individually sensible AI agents nevertheless produce collectively harmful outcomes? We study this question in a repeated commons dilemma with externalities. We introduce Creeping Trap, a controlled social simulation in which LLM agents repeatedly choose private extraction levels for immediate profit while jointly accumulating catastrophe risk and imposing harm on silent stakeholders. The environment includes analytically derived welfare and behavioral reference strategies, allowing individually sensible behavior and collective welfare to be evaluated simultaneously. Across nine contemporary LLMs and 990 episodes, including a 400-episode confirmatory study, simulated societies consistently over-extract; in the confirmatory study, 396 of 400 episodes produce negative aggregate welfare. This failure is not arbitrary irrationality: observed behavior lies near a broad low-regret region of local best response that even simple heuristics attain, indicating an institutional failure generated by individually sensible decisions. We further show that institutional design changes collective outcomes: surfacing stakeholder harm consistently reduces extraction, whereas shortening accountability horizons increases extraction and induces end-of-term defection. Finally, prompt wording substantially changes measured behavior, making prompt formulation a first-order experimental variable. Creeping Trap provides a controlled testbed for studying institutional interventions that improve collective outcomes in AI societies.

Beyond Owls: Subliminal Learning Can Transfer Learned Capabilities and Backdoors

In subliminal learning (SL), a teacher model passes on a trait to a student model by distillation on data semantically unrelated to the trait. So far, SL has been demonstrated for only a limited range of traits, including preferences for animals (e.g., owls) and malicious personas. These traits can also be elicited with simple prompts or with steering. Can SL transfer a wider range of traits, including more complex ones? If so, distillation might transfer subtle forms of misalignment (e.g., reward-seeking, scheming, and secret loyalties) without detection. To this end, we test whether SL can transfer a novel capability: predicting the outputs of a randomly initialized MLP. After distilling on unrelated text, the student achieves substantial performance on the task, while falling short of the teacher. We find that a directly optimized steering vector matches SL in distribution but generalizes worse out of distribution. Next, we test whether SL can transfer backdoors. We finetune the teacher to answer in French when the prompt contains a female name, then distill on number sequences containing neither names nor French. The student partially acquires the backdoor, responding in French on 23.5% of prompts with female names versus 0.0% with male names. Finally, we test whether SL can transfer a propensity to hack in an agentic chess environment. We finetune the student on number sequences from a steered hacker teacher. The student hacks in 58.3% of episodes, compared with 10.9% for the unfinetuned model. Thus, we show SL can transfer capabilities, backdoors, and hacking propensities. The amount of transfer is sensitive to the setup. In several experiments, it is made stronger by using logit distillation or by restricting LoRA to the attention layers.

The Majority Illusion in Social Networks

Social behaviors are often contagious, spreading through a population as individuals imitate the decisions and choices of others. A variety of global phenomena, from innovation adoption to the emergence of social norms and political movements, arise as a result of people following a simple local rule, such as copy what others are doing. However, individuals often lack global knowledge of the behaviors of others and must estimate them from the observations of their friends' behaviors. In some cases, the structure of the underlying social network can dramatically skew an individual's local observations, making a behavior appear far more common locally than it is globally. We trace the origins of this phenomenon, which we call "the majority illusion," to the friendship paradox in social networks. As a result of this paradox, a behavior that is globally rare may be systematically overrepresented in the local neighborhoods of many people, i.e., among their friends. Thus, the "majority illusion" may facilitate the spread of social contagions in networks and also explain why systematic biases in social perceptions, for example, of risky behavior, arise. Using synthetic and real-world networks, we explore how the "majority illusion" depends on network structure and develop a statistical model to calculate its magnitude in a network.

Mechanisms of social cognition

Social animals including humans share a range of social mechanisms that are automatic and implicit and enable learning by observation. Learning from others includes imitation of actions and mirroring of emotions. Learning about others, such as their group membership and reputation, is crucial for social interactions that depend on trust. For accurate prediction of others' changeable dispositions, mentalizing is required, i.e., tracking of intentions, desires, and beliefs. Implicit mentalizing is present in infants less than one year old as well as in some nonhuman species. Explicit mentalizing is a meta-cognitive process and enhances the ability to learn about the world through self-monitoring and reflection, and may be uniquely human. Meta-cognitive processes can also exert control over automatic behavior, for instance, when short-term gains oppose long-term aims or when selfish and prosocial interests collide. We suggest that they also underlie the ability to explicitly share experiences with other agents, as in reflective discussion and teaching. These are key in increasing the accuracy of the models of the world that we construct.

Social physics in the age of artificial intelligence

Artificial intelligence (AI) systems are rapidly becoming more capable, autonomous, and deeply embedded in social life. As humans increasingly interact, cooperate, and compete with AI, we move from purely human societies to hybrid human-AI societies whose collective dynamics cannot be captured by existing behavioural models alone. Drawing on evolutionary game theory, cultural evolution, and Large Language Models (LLMs) powered simulations, we argue that these developments open a new research agenda for social physics centred on the co-evolution of humans and machines. We outline six key research directions. First, modelling the evolutionary dynamics of social behaviours (e.g. cooperation, fairness, trust) in hybrid human-AI populations. Second, understanding machine culture: how AI systems generate, mediate, and select cultural traits. Third, analysing the co-evolution of language and behaviour when LLMs frame and participate in decisions. Fourth, studying the evolution of AI delegation: how responsibilities and control are negotiated between humans and machines. Fifth, formalising and comparing the distinct epistemic pipelines that generate human and AI behaviour. Sixth, modelling the co-evolution of AI development and regulation in a strategic ecosystem of firms, users, and institutions. Together, these directions define a programme for using social physics to anticipate and steer the societal impact of advanced AI.