Cumulative Culture and Social Learning in 2035: Scenarios, Signals, and Falsifiable Predictions

Whether platforms, cross-cultural samples, and AI-generated content will measurably change how humans copy, teach, and improve on each other's ideas by 2035 — stated as falsifiable bets, not forecasts of vibes.

A transmission-chain testing bench: the actual apparatus researchers use to measure whether a technique gains, degrades, or stalls as it passes from one learner to the next. — Image prompt and art direction by Brecht Corbeel; image generated to that direction.

Abstract

Cumulative culture — the human capacity to retain, recombine, and improve on what previous generations learned, rather than reinvent it each generation — rests on a small set of mechanisms: high-fidelity imitation, teaching, language, and population structure large enough to preserve rare skills against loss. This article lays out what the research actually shows about those mechanisms, then states three falsifiable scenarios for how the field and its subject will look by 2035: whether digital platforms measurably change transmission fidelity, whether social-learning research diversifies beyond WEIRD samples, and whether AI-generated content produces a documented, measurable shift in innovation and fidelity dynamics. Each scenario carries a horizon, stated assumptions, observable indicators, and an explicit condition under which it would be considered wrong.

Every human alive today benefits from techniques no single person invented and no single lifetime could reconstruct from scratch. The compass, the antibiotic, the suspension bridge, the sourdough starter with a lineage older than the country it is baked in — each is the residue of many separate minds adding small increments to what came before, and each increment surviving only because someone transmitted it faithfully enough for the next person to build on it rather than start over. Researchers call this capacity cumulative culture, and it is one of the few traits that separates human social learning from social learning in every other studied species, at least at the fidelity and scale required to keep a ratchet turning [9].

This article does three things. First, it lays out, plainly and without hedging into vibes, what the actual experimental and archaeological evidence says about the mechanisms that make cumulative culture possible: imitation, teaching, language, and population structure. Second, it separates what several vendors and commentators currently claim about AI’s effect on culture from what has actually been measured. Third, and this is the point of the piece, it states three falsifiable scenarios for where this field and its subject will stand by 2035 — each with a horizon, stated assumptions, an observable indicator, and an explicit condition under which the scenario would be considered wrong. A forecast without a disconfirmation condition is not a prediction; it is a mood.

What “cumulative” actually requires

Cumulative culture is not the same thing as social learning in general. Many animals learn from each other — chimpanzees crack nuts with stones they have watched others use, and some bird populations sustain local song dialects across generations. What is rare, and arguably unique to humans at the scale it occurs, is a ratchet: each generation’s modification is retained with enough fidelity that the next generation can add to it rather than merely repeat or slowly degrade it [9].

The clearest experimental demonstration of what makes the ratchet turn comes from Dean and colleagues’ 2012 study comparing children, chimpanzees, and capuchin monkeys on a three-level puzzlebox designed so that reaching the top level required building on a solution to the level below it. Three- and four-year-old children progressed through the levels; almost none of the chimpanzees or capuchins reached the highest level. The children’s success tracked a specific package of behaviors: imitation of the exact observed action (not just the goal), teaching — including verbal instruction — directed at other members of their group, and prosocial sharing of the rewards obtained. None of the three elements appeared reliably in the non-human primates [1]. This is a finding about mechanism, not a general claim of human superiority: it identifies which specific behaviors were present in the group that ratcheted and absent in the groups that did not.

Laboratory transmission-chain experiments extend this by isolating the ratchet under controlled conditions. Caldwell and Millen built physical microsocieties — small groups organized into chains of overlapping “generations” — around simple construction tasks: building the tallest possible tower from sticks and modeling clay, or the best-flying paper airplane from a single sheet. Performance improved across successive chain generations in ways that could not be explained by individual trial-and-error alone; participants who could observe a predecessor’s attempt reliably out-performed those working from scratch, and improvement compounded across links in the chain [2]. This is the experimental scaffolding underneath the diffuse popular claim that “humans stand on the shoulders of giants” — it is a measured, replicable effect with a known failure mode: break the chain, or degrade fidelity below some threshold, and the tower gets no taller.

Figure 1. Cumulative culture is legible only across a sequence: each solution must retain the previous one's gain before adding its own, or the chain resets to scratch. — Image prompt and art direction by Brecht Corbeel; image generated to that direction.

Teaching and imitation are not interchangeable

A common simplification treats “social learning” as a single dial that runs from none to a lot. The evidence argues against that. Legare and Nielsen’s synthesis of developmental work distinguishes two engines that operate on different rules depending on what is being learned. For instrumental tasks — where the goal is a functional outcome, like getting a toy out of a box — children imitate closely at first and shift toward innovation and individual problem-solving as they gain experience, exploring variants that might work better. For conventional tasks — where the goal is fidelity to a social norm, like following a game’s arbitrary rules — imitation stays high regardless of experience, and innovation stays low throughout [5]. This is an important qualifier on any claim about “transmission fidelity” as a single number: fidelity is task-dependent, and a platform or technology that raises fidelity for one kind of content (a recipe, a dance) may leave it untouched or even suppress it for another (a tool’s functional design).

Population size and connectivity matter independently of any individual’s learning strategy. Kline and Boyd’s analysis of Oceanic island societies at early European contact found that islands with larger, better-connected populations sustained more complex marine-foraging toolkits than small, isolated islands — a test made possible because contact rates between islands could be estimated directly, unlike in continental settings [3]. The complementary case is loss rather than complexity: Henrich’s analysis of Tasmania, cut off from mainland Australia by rising sea levels roughly twelve thousand years ago, documents the loss of a series of technologies — including bone tools, cold-weather clothing, and several kinds of hafted and thrown weapons — over the subsequent millennia, in a population that had not become less skilled or less intelligent but had become demographically too small and too isolated to reliably reproduce and transmit its own existing toolkit against ordinary transmission error [4]. Read together, these two findings support a specific claim: cumulative culture is not guaranteed by having culture at all; it depends on a population large and connected enough that transmission error does not outpace innovation. This is the empirical basis, not a metaphor, for asking what happens when a species radically changes the size and structure of the population any given idea can reach — which is exactly what digital platforms, and now AI-generated content, do.

Figure 2. Larger, better-connected populations sustain more complex toolkits than small isolated ones — a pattern read from real technological inventories, not assumed. — Image prompt and art direction by Brecht Corbeel; image generated to that direction.

What is fact, vendor claim, and unmeasured hope

Before turning to scenarios, it is worth separating categories of claim that get blended constantly in coverage of “AI and culture.”

Fact, peer-reviewed: Shumailov and colleagues demonstrated, in controlled model-training experiments (Gaussian models, variational autoencoders, Gaussian mixture models, and a language model), that training a generative model recursively on its own or its predecessors’ outputs, without a floor of fresh unmodified data, causes measurable “model collapse” — early-stage drift from the true data distribution followed by permanent loss of low-frequency (“tail”) variation, converging toward a narrower, more repetitive output distribution across generations [8]. This is a fact about machine-learning systems under specific recursive-training conditions, established in a controlled setting.

Vendor and commentator claim, not yet established as population-level fact: Various AI companies and commentators have asserted or implied that widespread AI-generated content circulating in public discourse will produce something analogous to model collapse in human culture — a narrowing of the range of ideas, phrasing, or styles that humans encounter and reproduce. This is a plausible extrapolation from the machine-learning result, and Acerbi’s broader analysis of digital cultural transmission gives reasons the analogy might hold — platforms already reshape which cultural variants get seen and copied through engagement-driven curation, independent of any AI content in the mix [7]. But extrapolating a training-time effect measured in machine-learning pipelines to human cultural transmission, which has different error-correction mechanisms (deliberate teaching, social correction, institutions, professional gatekeeping in many domains) than an unsupervised training loop, is an analogy, not a demonstrated finding. No published study as of this writing has measured a population-level narrowing of human cultural output attributable to AI-generated content in the transmission pool at the fidelity Shumailov’s team measured in models. Treat it as an open empirical question, addressed directly in Scenario 3 below.

Genuine, longstanding methodological problem, independent of AI: Henrich, Heine, and Norenzayan’s widely cited analysis found that a large share of foundational claims in psychology and behavioral science, including much of the social-learning literature that underpins the mechanisms described above, are drawn from WEIRD populations — Western, Educated, Industrialized, Rich, Democratic societies — which are statistical outliers on many of the measured dimensions of human psychology and behavior relative to the rest of the species [6]. This is not a claim about AI at all; it predates the current wave of AI-and-culture discourse by well over a decade, and it matters here because any confident claim about how “human cumulative culture” in general responds to platforms or AI is only as strong as the diversity of the populations it was measured in.

Figure 3. Most social-learning findings still come from a narrow slice of research sites; whether the sampled population diversifies by 2035 is an open, checkable question. — Image prompt and art direction by Brecht Corbeel; image generated to that direction.

Scenario 1: platform-mediated transmission fidelity

Horizon: by 2035 (roughly nine years from now).

Claim: Large-scale digital platforms will have produced a measurable, published, population-level change in cumulative-culture transmission fidelity — defined as the rate at which a specific technique, recipe, or convention is reproduced with functionally important detail intact across sequential “generations” of learners (analogous to the transmission-chain paradigm, but observed in the wild through platform data such as recipe-remix histories, open-source code forks, or tutorial-reproduction video pairs) — relative to a comparable pre-platform or lower-platform-exposure baseline.

Assumptions: That researchers gain some form of access to longitudinal platform data sufficient to trace content lineages (forks, remixes, stitched videos, recipe adaptations) at scale; that a comparable baseline population or time period exists for contrast; that “functionally important detail” can be operationalized similarly to how transmission-chain experiments already operationalize task success [2].

Observable indicator: A peer-reviewed study, published in a cultural-evolution, cognitive-science, or computational social science venue, reporting a statistically significant difference in transmission-fidelity metrics between a high-platform-exposure population or content category and a lower-exposure comparison, using a design that controls for the task-type distinction (instrumental versus conventional) that Legare and Nielsen’s work shows already produces different fidelity baselines [5].

Figure 5. Teaching, not only imitation, appears to be what lets a solution improve across a chain rather than merely survive it. — Image prompt and art direction by Brecht Corbeel; image generated to that direction.

Disconfirmation condition: This scenario is wrong if, by 2035, either no such study exists despite adequate data access, or multiple well-powered studies exist and find no consistent direction of effect — i.e., fidelity changes are present but bidirectional, small, and inconsistent across content domains, which would indicate platforms reshape what circulates far more than they reshape how faithfully it is copied once selected.

Scenario 2: beyond WEIRD samples

Horizon: by 2035.

Claim: Cross-cultural social-learning research will have measurably diversified its sampled populations relative to the 2010–2020 baseline documented by Henrich, Heine, and Norenzayan [6], such that a meaningful share of new foundational findings in the field are generated or replicated outside WEIRD populations, including in field settings analogous to the population-scale contrasts Kline and Boyd used in Oceania [3].

Assumptions: That funding, institutional access, and researcher training pipelines in non-WEIRD regions continue on something like their current trajectory rather than contracting; that “meaningful share” can be operationalized as a tracked proportion of sample origin in major social-learning and cultural-evolution journals, a metric already partially tracked by methodological review papers in this literature.

Observable indicator: A follow-up bibliometric review, comparable in method to the original WEIRD critique, finding a materially higher proportion of non-WEIRD-sampled foundational studies in top cultural-evolution and developmental-psychology venues in the 2030–2035 window than in the 2010–2020 window it critiqued.

Disconfirmation condition: This scenario is wrong if a comparable bibliometric follow-up finds the sample composition essentially unchanged, or changed only in token, low-weight ways (a handful of high-profile non-WEIRD studies cited disproportionately without a broad base beneath them) — which would indicate the field absorbed the 2010 critique rhetorically without changing its underlying research infrastructure.

Scenario 3: AI-mediated cultural transmission produces a documented shift

Horizon: by 2035.

Claim: The presence of AI-generated content in the general cultural transmission pool (text, images, tutorials, code, music, and similar culturally transmitted content encountered by human learners) will have produced a documented, measurable shift in either innovation rate or fidelity dynamics in human cultural transmission — analogous in kind, though not necessarily in magnitude or mechanism, to the model-collapse effect Shumailov and colleagues measured in recursively trained generative models [8] — as opposed to remaining an untested analogy as it is at the time of writing.

Assumptions: That the analogy between machine training-time collapse and human cultural transmission is even measurable with available methods; that researchers can distinguish an AI-content effect from the platform-curation effect already described in Scenario 1 and by Acerbi [7], which is a nontrivial confound since both push in a plausibly similar direction (narrowing of variety); that a suitable pre-AI baseline population or era remains available for comparison as AI-generated content becomes more pervasive.

Observable indicator: A peer-reviewed study measuring, in a human population, a change in either (a) the rate of genuine downstream innovation building on a piece of content, or (b) the survival rate of rare/tail variants of a cultural item (recipe variations, phrasings, solution types) across sequential reproduction — attributable specifically to AI-generated-content exposure rather than to platform curation alone, using a design capable of separating the two mechanisms (for example, comparing content ecosystems with differing measured AI-content saturation at a similar level of platform curation).

Disconfirmation condition: This scenario is wrong if, by 2035, either the confound between AI-content exposure and platform curation proves methodologically insurmountable (no study manages to separate them, and the field explicitly concludes this), or well-designed studies that do separate the two mechanisms find no measurable AI-attributable effect on human transmission fidelity or innovation rate distinct from ordinary platform dynamics — which would indicate the model-collapse analogy, however intuitive, does not transfer from training loops to human social learning.

Figure 4. Whether large digital platforms are quietly changing transmission fidelity, and whether AI-generated content shifts what gets copied, are the two open empirical questions this decade's data will settle. — Image prompt and art direction by Brecht Corbeel; image generated to that direction.

Why the disconfirmation conditions matter more than the scenarios

It would be easy to write a version of this article that simply asserted platforms are homogenizing culture and AI content will accelerate it — that is the ambient narrative, and Shumailov’s result gives it surface plausibility. It would be equally easy to write the opposite, asserting that human transmission has enough built-in error correction (teaching, social sanction, professional gatekeeping in many domains) that none of this matters. Both would be predictions with no way to be wrong, which makes them unfalsifiable and therefore not useful as forecasts, only as priors. The value of stating an explicit disconfirmation condition for each scenario is that it forces the claim to specify, in advance, what kind of published evidence would force abandoning it — and it makes clear that the honest current answer to all three questions is: not yet measured, actively contested, worth checking again in the next several years rather than a confident yes or no in either direction.

What the existing evidence base does support, without extrapolation, is that cumulative culture has always been contingent on population structure and transmission fidelity rather than guaranteed by having culture at all [3, 4], that the specific behaviors which make the ratchet turn are identifiable and have been isolated experimentally rather than merely inferred [1, 2], that the mechanisms operate differently depending on whether a norm or a function is being learned [5], that the empirical base underlying all of this is narrower in population coverage than its confident general claims imply [6], and that at least one directly analogous mechanism for cultural narrowing has been rigorously demonstrated in machine systems, even though its transfer to human populations remains unestablished [8, 7, 9]. Whether 2035 confirms, disconfirms, or simply leaves each of these three scenarios still open is, by design, a question this article commits to being wrong about in a checkable way.

Sources

1. Lewis G. Dean, Rachel L. Kendal, Steven J. Schapiro, Bernard Thierry, and Kevin N. Laland. Identification of the Social and Cognitive Processes Underlying Human Cumulative Culture. Science (2012). DOI: 10.1126/science.1213969.
2. Christine A. Caldwell and Ailsa E. Millen. Experimental models for testing hypotheses about cumulative cultural evolution. Evolution and Human Behavior (2008). DOI: 10.1016/j.evolhumbehav.2007.12.001.
3. Michelle A. Kline and Robert Boyd. Population size predicts technological complexity in Oceania. Proceedings of the Royal Society B (2010). DOI: 10.1098/rspb.2010.0452.
4. Joseph Henrich. Demography and Cultural Evolution: How Adaptive Cultural Processes Can Produce Maladaptive Losses—The Tasmanian Case. American Antiquity (2004).
5. Cristine H. Legare and Mark Nielsen. Imitation and Innovation: The Dual Engines of Cultural Learning. Trends in Cognitive Sciences (2015). DOI: 10.1016/j.tics.2015.08.005.
6. Joseph Henrich, Steven J. Heine, and Ara Norenzayan. The weirdest people in the world?. Behavioral and Brain Sciences (2010). DOI: 10.1017/S0140525X0999152X.
7. Alberto Acerbi. Cultural Evolution in the Digital Age. Oxford University Press (2019). DOI: 10.1093/oso/9780198835943.001.0001.
8. Ilia Shumailov, Zakhar Shumaylov, Yiren Zhao, Nicolas Papernot, Ross Anderson, and Yarin Gal. AI models collapse when trained on recursively generated data. Nature (2024). DOI: 10.1038/s41586-024-07566-y.
9. Alex Mesoudi and Alex Thornton. What is cumulative cultural evolution?. Proceedings of the Royal Society B (2018). DOI: 10.1098/rspb.2018.0712.

Originally published at https://absolutedigitalpublishers.com/articles/cumulative-culture-and-social-learning-in-2035-scenarios-signals-and-falsifiable-predictions.