Cultural Knowledge and Innovation: Insights from a New Empirical Study

How a large-scale study of human innovation converges with a conceptual model of social norms as information and what that convergence means for the science of human behavior.

---

Few human capacities are as consequential, or as poorly explained, as our ability to invent. We accumulate ideas and technologies across generations, building on what came before in a way no other species manages.

The standard account of how this happens leans heavily on social learning: we copy, teach, and transmit, and good ideas spread and stick.^{[5, 6, 7]} But that account has always left a conspicuous gap. Transmission explains how an innovation propagates once it exists. It says very little about where the innovation came from in the first place about the cognitive act of generation itself.

A new study in the Proceedings of the National Academy of Sciences by Anil Yaman, Shen Tian, and Björn Lindström, of the Karolinska Institutet and Vrije Universiteit Amsterdam, takes direct aim at that gap.^[1] Their answer is both intuitive and, until now, surprisingly under-tested. They showed, rigorously and quite conclusively, that innovation is guided by what we already know. Specifically, it is guided by semantic knowledge — our structured internal map of how concepts, objects, and their functions relate to one another.

This is an important contribution in its own right, but also speaks directly to a body of work close to home. For more than a decade, Dr. Jennifer Loughmiller-Cardinal and I have been developing a conceptual model of social norms and culture. The Cardinal Research Institute was created to advance that work and to foster the kind of interdisciplinary research that complements and extends it.

The PNAS findings map onto our model's predictions with striking precision. First, we'll look at why their findings are so important to social scientists. After that, we'll discuss its implications for our model.

Innovation Leverages Cultural Knowledge

Classical models of cultural evolution have often treated innovation as a kind of random variation, the cultural analogue of genetic mutation, a blind generator of novelty that selection then filters.^{[5, 8]} It is a useful simplification, but the authors argue it is insufficient to explain the open-ended, cumulative character of human invention. People do not explore the space of possible combinations at random. They explore it intelligently.

To test this, Yaman and colleagues paired an agent-based model with a large behavioral experiment of 1,243 participants.^[1] In the experiment, people played a discovery game in which the goal was to create new "innovations" by combining items. Some worked with familiar objects (rocks, branches, and the like) whose real-world properties and uses are richly understood. Others performed an identical task using abstract symbols stripped of any such meaning.

The contrast was stark. Participants equipped with semantic knowledge directed their exploration toward meaningful, promising combinations, succeeded more often, and generalized from earlier discoveries to new ones. Participants without it did not simply do a little worse — they performed no better than chance, falling back on shallow, undirected exploration even when social information from others was available to them.

When semantic knowledge and social learning were combined, however, the two interacted synergistically. They found that groups with access to both produced roughly twice as many unique innovations as groups relying on social learning alone.^[1]

The headline, then, is twofold. First, the generative engine of innovation is not noise. It is structured prior knowledge doing cognitive work. Second, that knowledge does not act in isolation but it multiplies the value of social learning, rather than merely supplementing it.

Why it matters, on its own terms

It is worth reflecting on the independent significance of this before connecting it to anything else.

For decades, the cultural-evolution literature has been notably stronger on transmission than on generation. The mathematical and ethnographic machinery of social learning, who copies whom, when copying beats individual trial-and-error, how fidelity enables cumulative culture, is well developed.^{[5, 6, 8]}

Alan Rogers' classic paradox even showed that social learning alone confers no adaptive advantage at equilibrium without some individual learning to keep information current,^[9] and Cecilia Heyes has argued that the very mechanisms of social learning are themselves culturally constructed "cognitive gadgets."^[10] Yet, across this body of work, the moment of invention itself has tended to be black-boxed as random variation.

By giving that black box an empirical interior and by showing experimentally that emptying it (the abstract-symbol condition) drops performance to chance. Yaman, Tian, and Lindström, supply something the field has needed: a cognitive foundation for the generative side of cumulative culture. It is a genuine advance in our understanding of why human innovation is as open-ended as it is, and it deserves to be read on those terms regardless of any particular theoretical commitments.

Where this meets our work

Here is where the paper becomes, for us, more than an interesting result. At the Cardinal Research Institute our mission is to understand the natural system of information exchange, comparison, and updating that underlies human behavior, what we call "the science of human being." Over a series of co-authored papers, we have developed a conceptual model of social norms and culture built entirely on that premise and it is that model the new study so closely echoes.

We begin by reconsidering what a social norm actually is.^[2] Rather than treating norms as informal rules or strategies that constrain behavior (the dominant framing across sociology, behavioral economics, and game theory^[11]) we treat them as information. A norm, in this view, is not a rule we obey but a mechanism for identifying when a rule might be needed and what information is relevant to a situation.

Norms, in other words, are how a community captures, compares, and retains the knowledge required to keep its environment predictable and to lower the cognitive cost of acting within it.^[2]

In our second paper, we describe norms as the leading edge of a collective search across an "information landscape"^[3] deliberately analogous to the fitness landscape of evolutionary biology.^[15] This information landscape is defined as the mutual information realized between an environment and the mental representations that agents form of it. This landscape is at once internal and external, individual and collective. Internally, it is the beliefs, expectations, and cognitive capacities of a person. Externally, it captures the social signals, cultural materials, and curated norms through which information is shared and interpreted.

Individuals run autonomous searches across it, sampling possible solutions while the community aggregates and normalizes those samples, updating the expectations that shape subsequent searches. Crucially, individual and collective optimization happen simultaneously and are entangled, with each continually reshaping the constraints and resources available to the other.

The mathematics implied here is that of predictive, surprise-minimizing inference. Somewhat along the lines of the Bayesian and free-energy framing of cognition associated with Karl Friston and Andy Clark^{[13, 14]} but scaled up from the individual brain to the distributed network of a community. This is loosely grounded in Claude Shannon's information theory, but enacted through the biological principle of allostasis, or stability through change, and homeostatic regulation.^{[12, 16]}

Our third paper extends the same logic to culture itself, defining it not as a collection of beliefs or traditions but as the configuration of "structural moments" of social information: norms (the searching edge), normativity (the convergence of that search toward stable expectations), and institutions (the curated, long-term repositories of validated knowledge).^[4]

Curated knowledge, in this account, becomes organized into semantically bounded domains (conventions, roles, rituals, protocols) and it is precisely this categorical and semantic structure that allows a community to extrapolate from limited experience into unfamiliar situations.^{[3, 4]}

Semantic knowledge as an internalized information landscape

Read against that framework, the PNAS results are less a surprise than a confirmation. The connections are direct, and the tightest of them is also the most fundamental, so it is worth stating precisely. What Yaman and colleagues call semantic knowledge, we would recognize as the individual, internalized face of what we call the information landscape.

Consider how each construct is defined. The PNAS authors describe semantic knowledge as the structured, generalizable associations linking concepts and objects to their properties and typical functions in their phrase, our internal map of how concepts relate.^[1] We define the information landscape as the mutual information realized between an environment and the mental representations an agent forms of it: a topology of belief, experience, and context across which individuals and communities search for useful, meaningful information.^{[3, 4]}

Strip the technical language from either definition and the same object appears, a structured representation that tracks the regularities of the world and tells an agent which possibilities are worth pursuing. The correspondence holds at three specific points.

It is mutual information between mind and world. Semantic knowledge works in the experiment precisely because familiar objects (rocks, branches) carry real associations with their properties and uses, while abstract symbols carry none. That is our definition of the information landscape almost verbatim: the mutual information between an environment and a mind's representation of it. The abstract-symbol condition is, in our terms, a deliberately flattened landscape (a search space drained of topology) and so, unsurprisingly, search across it reverts to chance.^{[1, 3]}

It is at once individual and collective, and recursive. The PNAS authors are explicit that semantic knowledge is not merely a private possession: it is socially transmitted, shaped by prior generations' innovations, and refined over time; at once an input to innovation and an output of it.^[1] This is exactly our claim that the information landscape is both internal and external, both individual and collective, and continuously regenerated through the entanglement of the two.^[4]

Each person's semantic knowledge is a sampled cross-section of the collective landscape; social transmission is how those cross-sections are reconciled into a shared one; and the curated result becomes the prior structure for the next round of search — our loop of norms, normativity, and institutions.^{[3, 4]}

It is the structure that search runs over. In our model, norms are the leading edge of a search across the information landscape, and the landscape supplies the contours that make that search tractable.^[3] In the experiment, semantic knowledge supplies exactly those contours — directing exploration toward meaningful regions of the combinatorial space and enabling generalization from one discovery to the next.^[1]

Semantic knowledge, in short, is what an individual searches with; the information landscape is what a community searches over. They are the same structure seen from the two ends of the individual–collective relationship.

Seen this way, the PNAS experiment reads as an unusually clean isolation of a single variable in our framework: hold the social machinery roughly constant, vary whether agents possess internalized landscape structure, and watch what happens to collective innovation. The answer — that without it, even socially available information cannot be exploited — is our model's central prediction rendered as a laboratory result.

And the broader pattern holds

Two further correspondences round out the picture:

• Guided search, not random variation. The single most important claim in the PNAS paper (that innovation is structured exploration rather than blind variation) is the founding premise of our model. Both reject the "random mutation" simplification, and both replace it with a search guided by structured prior information.

The experiment shows individual search collapsing to chance the moment that guiding structure is removed;^[1] our model says, in effect, that there is nothing left to search over once the structure is gone.^[3]

• Synergy as entanglement. Perhaps our most distinctive commitment is that individual and collective information processing are not separate stages but entangled, dual optimizations in which each side amplifies the other.^[3] The experimental finding that semantic knowledge and social learning interact synergistically (roughly doubling unique innovations relative to social learning alone^[1]) is a clean empirical signature of that entanglement. Neither input is sufficient by itself; the value lies in the interaction.

A note of appropriate care, the PNAS study operationalizes semantic knowledge at the scale of one person's object–function associations in a controlled task, while our model concerns the collectively curated structure of norms and institutions across a population. These are not identical objects.

But, as the semantic argument above suggests, they are plausibly the same process viewed at two scales; the individual's structured priors and the community's curated landscape mutually constituting one another, exactly as our entanglement claim requires. That a tightly controlled cognitive experiment and a broad theory of culture should meet in the middle, from such different starting points, is what makes the convergence compelling rather than circular.

Why this matters for social research and practice

The deeper significance, from where I sit, is methodological. Social science has long been criticized (sometimes fairly) for describing culture far better than it explains it.^{[4, 11]}

Definitions of "culture" and "social norm" proliferate without consensus on the underlying mechanisms, a problem at least as old as Kroeber and Kluckhohn's mid-century catalogue of competing definitions.^[17] Results like the PNAS study matter because they push the field from description toward mechanism and from cataloguing what people do toward explaining why the doing takes the shape it does.

That shift is not merely academic. If norms and culture are best understood as systems for structuring and searching information, then the practical levers change. Designing an effective educational sequence, a public-health communication, a policy intervention, or a socio-technical platform becomes a question of how to build, curate, and align the semantic structure people actually reason with — not how to push behavior with incentives bolted onto an unexamined black box of preferences.^[3]

Interventions that ignore the information architecture beneath behavior tend to misfire, as a generation of inconsistent "nudge" results and biased algorithmic systems has shown.^[3] Interventions that respect it have a far better chance of helping people, organizations, and institutions adapt.

This is the work we built the Cardinal Research Institute to do, and it is why a study about people combining rocks and branches in an online game is, to us, anything but trivial. It is a small, rigorous window onto the largest question in the social sciences — what culture is, and what it does.

Congratulations, and looking forward...

Empirical demonstration and theoretical explanation rarely arrive from such independent directions — a cognitive experiment out of Stockholm and Amsterdam, and the model of norms and information Dr. Loughmiller-Cardinal and I have been building in New York and meet so cleanly in the middle. When they do, both are strengthened. The theory gains an empirical anchor; the experiment gains a wider frame within which its result is not an isolated finding but a predicted one.

Congratulations to Anil Yaman, Shen Tian, and Björn Lindström on a significant contribution. We will be building on it, and we suspect many others will too.

Notes & References

1. Yaman, A., Tian, S., & Lindström, B. (2026). Semantic knowledge guides innovation and drives cultural evolution. Proceedings of the National Academy of Sciences. https://doi.org/10.1073/pnas.2530750123 (preprint: arXiv:2510.12837).

2. Loughmiller-Cardinal, J. A., & Cardinal, J. S. (2023). The Behavior of Information: A Reconsideration of Social Norms. Societies, 13(5), 111. https://doi.org/10.3390/soc13050111

3. Cardinal, J. S., & Loughmiller-Cardinal, J. A. (2024). Information, Entanglement, and Emergent Social Norms: Searching for 'Normal'. Societies, 14(11), 227. https://doi.org/10.3390/soc14110227

4. Cardinal, J. S., & Loughmiller-Cardinal, J. A. (2025). The Foundations of Culture and the Moments of Social Information. Heritage, 8(9), 386. https://doi.org/10.3390/heritage8090386

5. Boyd, R., & Richerson, P. J. (1985). Culture and the Evolutionary Process. University of Chicago Press.

6. Henrich, J. (2016). The Secret of Our Success: How Culture Is Driving Human Evolution, Domesticating Our Species, and Making Us Smarter. Princeton University Press.

7. Tomasello, M. (1999). The Cultural Origins of Human Cognition. Harvard University Press.

8. Mesoudi, A. (2011). Cultural Evolution: How Darwinian Theory Can Explain Human Culture and Synthesize the Social Sciences. University of Chicago Press.

9. Rogers, A. R. (1988). Does Biology Constrain Culture? American Anthropologist, 90(4), 819–831.

10. Heyes, C. (2018). Cognitive Gadgets: The Cultural Evolution of Thinking. Harvard University Press.

11. Bicchieri, C. (2006). The Grammar of Society: The Nature and Dynamics of Social Norms. Cambridge University Press.

12. Shannon, C. E. (1948). A Mathematical Theory of Communication. Bell System Technical Journal, 27, 379–423, 623–656.

13. Friston, K. (2010). The Free-Energy Principle: A Unified Brain Theory? Nature Reviews Neuroscience, 11(2), 127–138.

14. Clark, A. (2013). Whatever Next? Predictive Brains, Situated Agents, and the Future of Cognitive Science. Behavioral and Brain Sciences, 36(3), 181–204.

15. Wright, S. (1932). The Roles of Mutation, Inbreeding, Crossbreeding, and Selection in Evolution. Proceedings of the Sixth International Congress of Genetics, 1, 356–366.

16. Sterling, P. (2012). Allostasis: A Model of Predictive Regulation. Physiology & Behavior, 106(1), 5–15.

17. Kroeber, A. L., & Kluckhohn, C. (1952). Culture: A Critical Review of Concepts and Definitions. Peabody Museum of American Archaeology and Ethnology, Harvard University.