AI-Driven Cultural Evolution

Machine Discoveries Become Culture
Steering AI Creativity Beyond Cultural Reproduction
Collective Learning and Intelligence
Optimal Engagement–Diversity Tradeoffs in Social Media
Collective Incentives Improve Group Accuracy by Reducing Reliance on Shared News Sources

The ability of humans to create and disseminate culture is often credited as the single most important factor in our success as a species. Culture evolves as a function of the range of available traits (variation), the ways these traits can replicate (transmission), and the success they encounter (selection). Whenever there is a major transition in any of these three processes, cultural evolution is altered. For example, in the space of roughly 200 years (1300–1500), European culture was rocked by a succession of major transitions in variation, transmission, and selection. Through the tales of Marco Polo and other explorers, Europeans were exposed to a broad set of novel cultural traits, such as paper money and coal burning (variation); the invention of the printing press made it easy to replicate cultural traits with great fidelity (transmission); and the Black Death drastically changed the social and economic opportunities of commoners, reshuffling the value of cultural traits (selection).
Modern communication technologies have scaled some of the key factors that shape human cultural evolution, enabling the instant sharing of knowledge across the globe and the rapid self-organization of online communities of interest. CHM believes that a particular digital technology, namely AI, can substantially impact the process of cultural evolution, from recommender algorithms altering the flow of knowledge to AI agents becoming participants in the generation of culture itself, from music and visual art to scientific discoveries. The Center's long-term goal is to map the different ways in which AI will impact—or may already be shaping—human culture, and to establish a research agenda for behavioral scientists studying hybrid cultural evolution in the digital age.

Machine Discoveries Become Culture

AlphaGo's famously "alien" move foregrounded a central puzzle for human–machine culture: can humans do more than briefly imitate machine innovations, and instead learn, transmit, and sustain strategies that originate outside the human cognitive repertoire?

Although competitive domains such as Go and chess show that superhuman AI can reshape human behavior, it remains unclear under which conditions machine-originating innovations can be culturally transmitted and persist once machines are no longer present. To address this question, hybrid social learning was examined in a controlled laboratory setting using transmission chains in a sequential decision-making task designed to exploit a robust human bias toward myopic, loss-avoiding behavior. Participants repeatedly solved a network-based reward-search puzzle while observing solutions from a previous generation, allowing strategies to propagate through small populations over time. Introducing an algorithmic agent with complementary biases showed that algorithmic solutions reliably boosted the performance of immediately following participants, but this advantage proved fragile: copying fidelity was low, and machine-originated strategies rapidly disappeared over subsequent generations. These results show that machine innovations can locally improve human performance, yet human cognitive biases undermine their long-term cultural preservation.It was then investigated whether machine discoveries could be maintained at the population level when core conditions from cultural evolution theory are explicitly satisfied. Shifting the focus from short-term influence to long-term maintenance, the authors emphasized selective social learning, repeated transmission, and population dynamics. Human populations were exposed to strategies discovered by a model-free deep Q-learning algorithm whose computational exploration identified solutions exceedingly unlikely for humans to discover independently. By systematically varying discovery difficulty, transmission difficulty, and the recognizability of performance advantages, machine-discovered strategies were allowed to compete with entrenched human heuristics under realistic social learning dynamics.

Key References

Steering AI Creativity Beyond Cultural Reproduction

Can AI agents generate ideas that are both original and coherent, not merely random combinations?

While LLMs excel in well-defined tasks, they struggle in open-ended creative domains. When pushed toward novelty, they often lose coherence; when constrained to coherence, they reproduce culturally typical ideas. This tension reflects a deeper limitation: LLMs are strongly anchored to dominant cultural distributions learned from data, which restricts long-horizon creative exploration. Here, as AI systems are now actively participating in the loop of cultural evolution, understanding what kinds of open-ended cultural dynamics they can generate becomes critical. If such limitations are inherent to AI systems, then their integration into cultural feedback loops risks amplifying cultural convergence. Conversely, if mechanisms can be found to overcome this problem, AI systems may instead enable new feedback loops that actively diversify culture rather than homogenize it.

To this end, the Cultural Alien Sampler (CAS) was introduced as a concept-selection mechanism that explicitly separates conceptual coherence from cultural typicality. CAS uses two GPT-2 models fine-tuned on WikiArt concepts: (1) the Concept Coherence Model that scores whether concepts plausibly co-occur within artworks, and (2) the Cultural Context Model that estimates how typical those combinations are within an individual artist's body of work. By selecting concept combinations that are high in coherence but low in cultural typicality, CAS promotes ideas that are compositionally sound yet culturally alien.

In human evaluations (N = 100), CAS-guided agents significantly outperform GPT-4o and random baselines on both originality and harmony, and achieve performance comparable to human art students. Quantitative analyses show that CAS explores a broader conceptual space, exhibits lower repetition, and sustains long-range exploration without premature convergence, while GPT-based agents rapidly saturate and revisit familiar regions of the idea space. The results demonstrate that creative limitations of frontier LLMs are not merely due to model size, but to how cultural knowledge is sampled. CAS provides a computational framework for studying machine culture, showing how artificial agents can be designed to intentionally step outside learned cultural norms to unlock creative potential.

Key Reference

Collective Learning and Intelligence

Which factors facilitate success in human-AI collective learning and problem solving?

This line of work explores the various factors that determine the performance of teams as they make predictions, perform collaborative tasks, or collectively learn. A combination of simulation, online experiments, and descriptive data analysis is employed to explore these questions. The role of artificial intelligence in enhancing human collective intelligence is also examined, ranging from mediating social learning (e.g., recommender systems that rewire the social network) to enhancing sense-making and foresight (e.g., improving prediction by combining human and AI predictions).

A key question in collective learning is how skill diversity within teams interacts with the structure of their social network to shape performance. Using a computational model of collective learning, one project demonstrates that the effect of diversity is not uniformly beneficial but depends critically on both task complexity and network density. For simple tasks, diversity consistently hinders collective performance. However, for complex tasks, the picture is more nuanced: homogeneous groups outperform diverse ones when social networks are sparse, while diversity becomes advantageous in densely connected networks (Figure 4). These findings offer a mechanistic account of when and why diversity helps or hurts, and carry practical implications for team composition—suggesting that as connectivity increases in modern organizations and digital environments, leveraging diversity becomes increasingly important for tackling complex problems.

While collective intelligence has traditionally been understood through the lens of human-to-human social learning and innovation, intelligent algorithms—such as search engines, recommender systems, and large language models—now play a growing role in mediating the flow of information between people. This study extends a theoretical model of cumulative cultural evolution to incorporate a simplified form of algorithmic mediation alongside network-based social learning, examining how the two channels jointly affect the accumulation of complex cultural knowledge. The results show that algorithmic mediation has a significant impact on cultural accumulation, particularly in sparsely connected networks where direct social learning opportunities are limited. Crucially, the highest levels of cultural accumulation emerge when social learning and algorithmic mediation are combined, with the optimal balance between the two depending on network density. This work represents an early formal step toward understanding how human–AI interaction reshapes the collective processes that drive cultural and intellectual progress, offering a framework for generating testable hypotheses about the role of AI in augmenting human collective intelligence.

Key References

Optimal Engagement–Diversity Tradeoffs in Social Media

How much user engagement must a social media platform sacrifice in order to guarantee a given level of diversity in the content its users are exposed to?

This line of work investigates how algorithmic curation on social media shapes the information environment, focusing on the widely discussed but rarely quantified tension between maximizing engagement and ensuring exposure to diverse content. While the prior literature treats this tradeoff as an underlying assumption, no earlier theoretical framework jointly models the two within a single system. A combination of theoretical modeling and simulation experiments (based on empirical Twitter data) is employed to investigate the tradeoff: (i) as a worst-case guarantee, and (ii) as a practically achievable Pareto frontier, which can be understood as a set of best available trade-offs where improving one goal would require sacrificing another. The aim is to identify regimes where promoting diversity costs platforms little, since these are precisely the conditions under which platforms might plausibly choose to break up echo chambers (areas with low content diversity).

A central question is whether echo chambers are an unavoidable byproduct of engagement maximization. A stylized model is introduced in which social media users are modeled as nodes of a directed follower graph, tweets are partitioned into T types with user-specific retweet probabilities, and the platform controls an injection policy that exposes each user to additional content beyond what their followers retweet. Engagement is defined as expected retweets in the limiting state, while diversity is defined pessimistically as the minimum, across all users and all types, of the expected number of tweets of that type each user sees. The main theoretical result establishes a tight upper bound on the cost of guaranteeing a certain amount of diversity. Broadly speaking, when users are engaged in a wide variety of content types rather than narrowly fixated on one, diversity is cheap; importantly, this bound holds for arbitrary social graphs, showing that elaborate network dynamics do not worsen the tradeoff beyond the empty-graph baseline.

Complementing this worst-case analysis, an empirical study on Twitter datasets concerning gun control and abortion charts the actual Pareto frontier between engagement and diversity. The follower graph is reconstructed directly from the data, and tweet types are inferred by applying community detection to a co-occurrence network of the most frequent hashtags. Solving large-scale linear programs over these reconstructed instances reveals that the practical tradeoff is dramatically more favorable than the theoretical bound (Figure 5): a platform willing to sacrifice 5% of engagement secures roughly 12% of the maximum achievable diversity, and sacrificing 10% pushes this above 40% under higher-engagement settings, with the cost of diversity typically decreasing as overall retweet probabilities scale up. Together these findings suggest that breaking up echo chambers may be substantially less costly than commonly believed, providing a quantitative foundation for future work connecting diversity interventions to downstream effects on polarization.

Key Reference

Collective Incentives Improve Group Accuracy by Reducing Reliance on Shared News Sources

How do reward structures influence collective accuracy in online news consumption?

Research on online news consumption has long focused on improving people's ability to select accurate news sources. However, people who consume accurate information do not necessarily make an accurate collective. Reliable information aggregation and collective decision-making also benefit from independence between judgments. We tested whether collective incentives can improve collective accuracy by reducing judgment correlation, even when average individual accuracy remains unchanged. We define collective incentives as reward structures that tie an individual's payoff to the accuracy of their group's aggregated decision rather than their own personal accuracy. 

In an online experiment (N = 232), participants made forecasts after viewing realistic tweets from correlated or independent news sources (Figure 6). Shared tweets among group members were more accurate but fully correlated across group members; local tweets were less accurate but independent between participants. When rewarded for collective accuracy, participants used more independent local tweets, improving collective but not individual accuracy. Generalizing our findings, an exploratory analysis of public news datasets found a greater content correlation among larger publishers, modulated by accuracy. These findings highlight the overlooked importance of judgment independence alongside accuracy in creating thriving information ecologies online. Collective incentives may complement accuracy-based interventions by nudging news consumption towards diverse, local, and independent sources. We discuss how future research should implement collective incentives to improve collective decisions in real-world settings.

Key Reference