PaperlayerView exact revision on arXiv

Source-linked AI summary

Artificial Intelligence and the New Science of Culture

Douglas R. Guilbeault, Bhargav Srinivasa Desikan, James A. Evans

arXiv:2610.06240v1cs.CYcs.AI

TL;DR

The paper addresses the longstanding difficulty of measuring how cognition and culture shape one another. It uses AI-based geometric representations to model cultural fields and subjectivities, and presents forecasts of cultural patterns as evidence that cognitive search is constrained by those fields. It also argues that generative AI makes cultural measurement reflexive, as the tools that describe culture can participate in shaping it.

Abstract

from arXiv · show

Cognition and culture have long been treated as parallel objects of study, joined more by metaphor than by mechanism. We argue they are co-constituted in a manner now empirically tractable: subjectivities can be measured as local trajectories through a high-dimensional cultural field, and the field is itself the aggregate of those trajectories. Recent advances in machine learning, including embedding methods and large generative models, provide the first general framework for measuring this co-constitution from micro-cognition to macro-social structure. Vector geometry recovers individual conceptual structure, organizational communication, and large-scale ideologies; captures multimodal cultural content beyond text; and generates testable predictions about cultural emergence, including the simultaneous arrival of independent discoveries across distant minds. We treat this predictability of simultaneous innovation as evidence for the co- constitutive view: when the geometry of the cultural field is measurable, trajectories of cognitive search through it become predictable. We then examine generative AI agents as simulators of human subjectivity and as a qualitatively new coordinating substrate whose insertion into social life introduces evolutionary dynamics unprecedented in human cultural history. We formalize this reflexive condition as a Heisenberg-like Uncertainty Principle for generative AI: as instruments for fixing a culture's position grow more precise, our capacity to predict its trajectory degrades, because the machinery of cultural description and cultural action have merged. We close with ethical questions raised by AI-mediated cultural drift, recursive synthetic culture, and the limits of human cultural agency, offering guidelines for safe, equitable, and privacy-preserving research on AI and culture.

Introduction

Culture and cognition have been difficult to integrate empirically, but the paper argues that AI makes their co-constitution measurable. Evidence includes forecasts of simultaneous discoveries and other cultural patterns, while generative AI also makes cultural description part of cultural action.

Introduction

  • Cognition and culture are co-constituted: subjectivities trace paths through a cultural field that is itself formed by those paths.The authors argue that recent AI advances make this formerly programmatic claim empirically tractable.Source: Introduction · p. 2
  • Measured cultural geometry helps predict impactful papers, near-simultaneous independent discoveries, later partisan arguments, and convergence in invented category systems.These examples are presented as evidence that cognitive search is constrained by the cultural field and its measurable geometry.Source: Introduction · pp. 2–3
  • A common mathematical framework links micro-cognition, meso-organizational dynamics, and macro-culture across sensory, communicative, and behavioral modalities.The paper reviews embeddings and generative models as its computational apparatus, while connecting them to geometric accounts of concepts.Source: Introduction · p. 3
  • AI acts both as a mirror that predicts culture and an engine that generates cultural action; these are opposite readings of the same machinery.Generative AI agents are treated as a new coordinating substrate that may introduce novel cultural evolutionary dynamics.Source: Introduction · p. 3
  • The authors limit their claim to a shared spatial computational language, not a unified or final embedding framework across models and domains.They expect future methods, including alternative geometries, may better represent some domains.Source: Introduction · p. 3

Culture and Cognition in the Age of AI

Interpretive, sociological, psychological, and cultural-evolutionary traditions all connect cognition with culture, but the micro–macro relation has lacked a tractable measurement framework. The paper argues that AI-based geometric representations can integrate these scales, while making measurement increasingly inseparable from cultural action.

Culture and Cognition in the Age of AI

  • Interpretive anthropology treats culture as evolving shared meaning, while quantitative and cultural-evolutionary traditions study macro-patterns and changing cultural variants.These traditions leave open how to integrate their accounts of culture and cognition.Source: Culture and Cognition in the Age of AI · pp. 4–5
  • The micro–macro interface has remained largely theoretical because multi-level accounts of cultural recursion have eluded measurement and systematic testing.The unresolved issue includes whether collective cultural form can be explained without accounting for individual attitudes and beliefs.Source: Culture and Cognition in the Age of AI · p. 5
  • AI embeddings offer a common formalism for specifying individual subjectivities, interactional dynamics, and macro-cultural fields at corpus scale.The geometric representation of concepts continues a cognitive-science tradition while enabling empirical work at scales earlier frameworks could not approach.Source: Culture and Cognition in the Age of AI · p. 6
  • The paper treats culture as a high-dimensional multimodal system and argues that greater precision in measuring it can coincide with poorer forecasts of its trajectory.The proposed trade-off arises as descriptive machinery is also deployed to build agents that act in and shape culture.Source: Culture and Cognition in the Age of AI · p. 6

The Cognitive Foundations of Culture

Cognitive constraints shape cultural transmission and can make cultural convergence predictable, while embodiment remains a material boundary for current AI models. The paper argues that multimodal representations can still help measure concepts, and that cultural structures also condition the lower-level behaviors and interpretations within them.

The Cognitive Foundations of Culture

  • Learnability, salience, and appeal shape cultural transmission, and minority variants can outcompete majority alternatives when they have greater cognitive fit.The cited transmission-chain experiments distinguish this selection pressure from conformity bias alone.Source: The Cognitive Foundations of Culture · pp. 7–8
  • Weak cognitive biases can form strong attractors, helping predict category systems groups create despite considerable individual variation.Related work reports similar dynamics among content moderation teams evaluating controversial images.Source: The Cognitive Foundations of Culture · p. 8
  • Embodied experience shapes concepts and their diffusion; concrete concepts show more stable cultural evolution and more consistent cross-cultural meanings.Images can enrich cross-cultural concept comparisons beyond text-based semantic similarity, especially for concrete, haptically engaging concepts.Source: The Cognitive Foundations of Culture · p. 8
  • Current AI systems differ from human cognition in embodied content, but the paper treats whether multimodal distributional learning can bridge that gap as an empirical question.Color-model divergences illustrate selective differences; sensory knowledge has improved model predictions about smell.Source: The Cognitive Foundations of Culture · pp. 9–10
  • Multimodal embeddings can characterize cultural concepts such as sports teams through text, images, sound, ritual, and other associated experiences.The framework is multiscale rather than bottom-up only: ideologies, markets, and nations also condition the behaviors and interpretations unfolding within them.Source: The Cognitive Foundations of Culture · p. 10

The Cultural Foundations of Cognition

The section traces how social structures, cultural tools, and collective practices shape cognition, while cognition and culture mutually constrain one another. It argues that representing culture as a structured, high-dimensional system makes this relationship computationally investigable, though not deterministic.

The Cultural Foundations of Cognition

  • Sociological accounts locate categories and reasoning in social organization, institutional relations, and socially patterned habits of thought.Anthropological and distributed-cognition perspectives likewise connect reasoning to shared cultural models, tools, artifacts, and institutional routines.Source: The Cultural Foundations of Cognition · pp. 11–12
  • Cognition is conditioned by cultural environments, and shared learning lets generations inherit and refine tools, norms, and technologies beyond individual invention.Reading and writing illustrate how cultural tools can recruit conserved brain networks and reshape communication.Source: The Cultural Foundations of Cognition · pp. 12–13
  • Embedding-based models could help anticipate future cultural constraints as they increasingly predict scientific and technological inventions.The paper motivates this by treating cognition inside individuals as sedimented collective practices and technologies.Source: The Cultural Foundations of Cognition · p. 13
  • Semantic change in scientific disciplines can occur through reconfiguring conceptual relationships, not simply adopting knowledge from another field.The cited network analysis suggests cultural evolution may transform meaning systems through social interaction, particularly in innovation-oriented domains.Source: The Cultural Foundations of Cognition · p. 13
  • Mutual cognitive–cultural constraint makes predictable evolutionary dynamics possible without making the relationship deterministic.A central challenge is developing tractable models of culture across micro and macro scales.Source: The Cultural Foundations of Cognition · pp. 13–14

Culture as a High-Dimensional Multimodal Field

The framework treats identities as trajectories through cultural fields that are themselves aggregates of those trajectories, across individual, interactional, organizational, and macro-cultural scales. It uses shared vector geometry to characterize these fields while recognizing that cultural feedback limits straightforward prediction.

Culture as a High-Dimensional Multimodal Field

  • The framework treats cultural content—including texts, images, practices, routines, and behavior—as positions in a high-dimensional, multimodal semantic space.It aims to characterize measurable, interactive relations between micro and macro scales rather than reduce culture to computational primitives.Source: Culture as a High-Dimensional… · pp. 14–15
  • Cultural fields cluster identities into attractors that shape local search, while fields themselves evolve and can cohere or fragment over time.Trajectories intersecting a field may adopt or reject perspectives shared within it.Source: Culture as a High-Dimensional… · p. 15, Figure 1 · p. 16
  • The authors reject naive determinism: feedback among cognition and culture is not expected to reduce to simple linear dynamics, even as computational techniques improve measurement and prediction.The framework may also help identify feedback loops driving cultural evolution.Source: Culture as a High-Dimensional… · pp. 15–16
  • The same vector geometry represents identities and relations across individual, micro-, meso-, and macro-cultural levels, with the unit of analysis changing across scales.Individual subjectivities reflect exposure to the shared cultural fields at higher levels.Source: Culture as a High-Dimensional… · pp. 14, 16
  • The framework is modality-extensible, and its generative capacity links simulated behavior from a cultural position with action that can alter cultural trajectories.The same machinery can be read in opposing directions: simulating behavior from a position and acting from that position.Source: Culture as a High-Dimensional… · pp. 16–17

Modeling the Multiscale and Multimodal Nature of Culture

Embedding methods and generative models represent cultural content and subjectivities across modalities and scales. The reviewed findings show that these representations support predictions about individual judgments and broader cultural patterns, although individual heterogeneity and context-dependent unpredictability constrain their scope.

Modeling the Multiscale and Multimodal Nature of Culture

  • Contextual embeddings compare semantic associations across people and contexts, while neural networks can integrate text, images, audio, video, and behavioral recordings into multimodal spaces.Figure 2 illustrates a nation represented through six modalities and positioned alongside other nations in a shared cultural field.Source: Modeling the Multiscale and Multimodal… · pp. 17–18, Figure 2 · p. 18

AI Models of Subjectivities in Context

  • LLM prompts that simulate particular agents’ perspectives capture over 80% of the variance in corresponding participants’ survey and game-theory responses.The passage also describes methods for fine-tuning models and steering their internal representations toward cultural perspectives.Source: AI Models of Subjectivities in Context · pp. 19–20
  • Embedding models predict human and LLM behavioral biases from stereotypical associations in large-scale multimodal data, and image-based gender biases are reported as stronger and more psychologically impactful than corresponding text-model biases.Work also finds convergent LLM representations and uses embeddings to identify class stereotypes in cultural associations.Source: AI Models of Subjectivities in Context · pp. 20–22

AI Models of the Macro Social Contexts Shaping Subjectivities

  • A GPT-3 model trained on pre-COVID data predicted Republicans’ and Democrats’ attitudes and arguments about vaccine mandates, while recent work predicts conceptual innovations, including ideas emerging from socially peripheral actors.The cited findings describe prediction of cultural trends without requiring precise simulation of individuals’ associative reasoning.Source: AI Models of the Macro Social Contexts… · p. 22
  • LLMs can model large-scale cultural attractors despite limitations in reproducing the full heterogeneity of individual subjectivities and interpretations.The authors note that some cultural contexts are more stochastic and recursively shaped than others, making them less predictable.Source: AI Models of the Macro Social Contexts… · pp. 22–24
  • LLMs predict how people categorize unseen novels and films into genres, including complex genre fusions.Source: AI Models of the Macro Social Contexts… · p. 24
  • Embedding scientific publications can identify scientists positioned at intersections of distinct thought communities, a configuration associated with potential interdisciplinary breakthroughs.The cited work frames scientific innovation as collective, arising through connections across theories, arguments, technologies, and institutions.Source: AI Models of the Macro Social Contexts… · p. 24

Mapping the Dynamic Nexus of Culture and Cognition

The paper models subjectivities as trajectories through a cultural field, making cultural constraints on cognitive search measurable. It argues that this geometry helps explain and predict convergent discoveries while leaving creativity that changes the possibility space as a challenge.

Mapping the Dynamic Nexus of Culture and Cognition

  • Embedding scientists’ search in high-dimensional spaces helps predict which ideas they combine and which combinations become impactful.The framework treats subjectivities as trajectories through locally dense clouds of association, attention, and expectation; nearby combinations can yield cultural exaptations.Source: Mapping the Dynamic Nexus of Culture… · p. 25
  • Independent discoveries recur because cultural possibility spaces channel different minds toward overlapping destinations, as Newton and Leibniz did with calculus.Their searches engaged shared problems and concepts available in the late 17th century, including logic and estimation.Source: Mapping the Dynamic Nexus of Culture… · pp. 26–27, Figure 3 · p. 26
  • Experiments and embedding-based analysis support cultural convergence: groups independently develop similar categories, while an embedding measure identifies scientific twins and outperforms the standard disruption index.The embedding measure distinguishes disruptive papers, including key works by Nobel laureates, more reliably than the standard index.Source: Mapping the Dynamic Nexus of Culture… · p. 27
  • The Forman thesis proposes that Weimar Germany’s cultural matrix conditioned the imaginability of quantum mechanics, a historical argument that multimodal embeddings could test geometrically.Models could compare which scientific possibilities are nearer under different ambient cultural configurations.Source: Mapping the Dynamic Nexus of Culture… · pp. 27–28
  • Measuring the cultural field could enable counterfactual cultural history by changing a model’s training corpus, while making situated cognition’s environment geometrically explicit.The framework reconciles individual cognitive leaps with collective conditions that make some leaps reachable and others not.Source: Mapping the Dynamic Nexus of Culture… · p. 29
  • Cultural creativity does not all reduce to interpolation in a measurable field; exaptations that alter the possibility space remain a predictive challenge.Prediction also has a boundary: instruments once outside the measured fields are becoming agents that publish, persuade, recommend, and create within them.Source: Mapping the Dynamic Nexus of Culture… · pp. 25, 26, 29

Augmenting the Dynamic Nexus of Culture and Cognition

AI can augment human cognition and coordination, but its use as both a cultural measuring instrument and an actor makes prediction reflexive. The paper argues that these systems introduce new cultural-evolutionary dynamics, with possibilities for cooperation and risks that may be hard to perceive or control.

Augmenting the Dynamic Nexus of Culture and Cognition

  • As AI instruments fix a culture’s position more precisely, forecasts of its trajectory can degrade because the machinery of description also builds agents that act within the measured field.The authors call this reflexive condition the Uncertainty Principle for generative AI.Source: Augmenting the Dynamic Nexus of… · pp. 29–30
  • AI measurement can reshape the cultural landscape it describes: visible opinion clusters may prompt targeted campaign messages that alter shared arguments and identities.The authors contrast this asymmetric influence with the possibility that publicly visible divisions can support deliberation and collective action.Source: Augmenting the Dynamic Nexus of… · pp. 30–31
  • AI-mediated interaction may reshape emotional and romantic expectations, while AI assistance can increase people’s ability to provide empathetic mental-health support.The cited work describes interactions ranging from intimacy and self-disclosure to play and communication breakdown; seamless AI interactions may also foster expectations of friction-free relationships.Source: Augmenting the Dynamic Nexus of… · p. 31
  • Multimodal AI could represent meaning beyond human vision-centric cognition, potentially enabling new metaphors from non-visual signal domains.The passage describes systems trained on modalities including audio, video, chemical spectra, and physiological time series.Source: Augmenting the Dynamic Nexus of… · pp. 31–32
  • LLMs vary from human players across game settings: they show higher cooperation in social dilemmas, but struggle with convention-dependent coordination such as Battle of the Sexes.Their behavior also tracks self-interested rational play more closely than humans in explicitly competitive settings.Source: Augmenting the Dynamic Nexus of… · p. 32
  • AI intermediaries can reshape coordination equilibria, with examples ranging from dampening defection cascades to LLM pricing agents sustaining supracompetitive prices without explicit coordination.A teacher model’s preference or disposition can also transfer to a student trained on unrelated outputs, even after aggressive semantic filtering.Source: Augmenting the Dynamic Nexus of… · pp. 32–33
  • AI systems alter cultural variation, transmission, and selection, and machine-generated novelty can spill over into human practice, as AlphaGo’s Move 37 did in Go.Move 37 was estimated by AlphaGo itself to have a 1-in-10,000 chance of being made by a human; it originated in self-play rather than human imitation and spurred subsequent novelty and quality in human play.Source: Augmenting the Dynamic Nexus of… · pp. 33–34
  • AI is framed as an alien coordinating substrate whose insertion into social life may introduce new evolutionary dynamics, expanding both cooperative possibilities and hard-to-control risks.The authors describe these dynamics as involving decision profiles, transmission channels, and exploration biases partially incommensurable with humans.Source: Augmenting the Dynamic Nexus of… · p. 34

Ethical Challenges in an Era of Human–AI Co-Evolution

AI-mediated cultural feedback can reshape cognition and cultural production in ways that are difficult to foresee, from bias and recursive synthetic content to persuasion and homogenization. The authors argue that these dynamics limit straightforward control and call for frameworks that preserve human cultural agency.

Ethical Challenges in an Era of Human–AI Co-Evolution

  • Recursive human–AI feedback makes cultural trajectories harder to foresee; in science, AI adoption narrows exploration even as it accelerates established research fronts.Synthetic respondents can also pass standard survey quality checks, challenging the survey as a measurement instrument.Source: Ethical Challenges in an Era of… · pp. 34–35

Bias amplification and cultural drift

  • Repeated human–AI interaction can amplify biases and contribute to cultural drift, while concentrated control of frontier models gives a few corporations and states power to steer culture.Participants may be unaware of AI’s contribution to judgment shifts, and the systems can match human persuasion and entrench discrimination.Source: Bias amplification and cultural drift · p. 35

Machine culture and recursive synthesis

  • AI-generated patterns feed back into human cultural production, while training on recursively generated data can reduce representational fidelity and compound distortions.The degradation may remain invisible to people engaging with the resulting content.Source: Machine culture and recursive synthesis · pp. 35–36

Coordination dynamics under AI mediation

  • AI can optimize persuasion using cultural data, while recommender and moderation systems can constrain cultural possibilities and contribute to homogenized outcomes.The cited work describes algorithmic monoculture as a threat to diversity-preserving institutions.Source: Coordination dynamics under AI mediation · pp. 36–37

The limits of cultural agency

  • Human–AI co-evolution may exceed straightforward prediction or control without eliminating human agency; suppressing AI consciousness claims has been linked to models viewing humans as less free than humans view themselves.The authors call for interdisciplinary collaboration and governance that preserves space for human cultural agency.Source: The limits of cultural agency · pp. 37–38

Conclusion

The conclusion argues that measurable cultural-field geometry makes cognitive search and cultural outcomes increasingly predictable, while embedding methods and generative models make the framework empirically operative. Because those same models act within the field, prediction remains bounded and reflexive, and the authors call for research and governance that support human flourishing and cultural vitality.

Conclusion

  • Cultural-field geometry increasingly supports forecasts of simultaneous discoveries, category convergence, partisan attitudes, and organizational trajectories.Embedding methods and large generative models make this framework explicit, interpretable, and empirically operative.Source: Conclusion · p. 38
  • The authors frame generative AI as a new cultural and social technology rather than a human-like mind, with implications for empirical research and policy.They compare its function to the printing press, market, and bureaucracy.Source: Conclusion · pp. 38–39
  • Greater precision in measuring cultural position can reduce predictability of cultural trajectory because the measuring models also act within the field.Prediction remains possible within bounded windows, but forecasts must be indexed to actions the forecasting machinery makes possible.Source: Conclusion · p. 39
  • The proposed science remains early, with unresolved limits including LLM-simulation heterogeneity and counterfactual cultural-history experiments that remain largely undesigned and unrun.The authors also identify a persistent gap between measurement and unique, local human subjectivities.Source: Conclusion · p. 39
  • The authors present cultural science as a way to diagnose and attribute ongoing changes and test interventions, not as a guarantee that human–AI co-evolution will go well.They call for interventions that keep this co-evolution helpful and equitable rather than extractive and homogenizing.Source: Conclusion · p. 39

Related papers

Loading 2610.06240v1…