The tragedy of the cognitive commons: collective intelligence beyond AI-induced knowledge collapse

Abstract: In a recent dynamic model by Acemoglu, Kong and Ozdaglar (2026a) agentic AI can cause a self-reinforcing deterioration of humanity's common knowledge base in what they call knowledge collapse. The model is based on a natural complementarity between the cumulative general knowledge of humans and locally generated context-specific knowledge, and on a learning externality that means that we all contribute to the private signal and the thin public signal that feeds the collective stock. If agentic AI can substitute for the private signal but not rebuild the public signal, and when human effort is sufficiently elastic, we can reach a low-knowledge equilibrium. This paper offers a measured appraisal of the model. It assesses the model in terms of what the popular summaries say is common knowledge, the world literature against knowledge commons and model collapse, and the partial empirical evidence, such as a 25% decline in public knowledge sharing on Stack Overflow. It then presents five structural criticisms of the model: its fixed knowledge taxonomy, the assumption that AI cannot provide general knowledge, the unmeasured effort elasticity parameter, and the political futility of the model to be implemented. We argue that the model's value lies not in any guaranteed catastrophe but in identifying a credible negative externality, the tragedy of the cognitive commons, whose remedy is a better aggregation of human knowledge. The paper outlines a research agenda for the study of effort elasticity and knowledge commons.

• The paper reframes AI-driven knowledge collapse as a conditional tragedy of the cognitive commons, showing that high effort elasticity, knowledge depreciation, and weak public-signal productivity can create multiple equilibria and declining collective knowledge.
• The paper argues that evidence such as reduced Stack Overflow activity and cognitive-debt findings supports local mechanisms but cannot yet establish societal-scale collapse, especially because reasoning may shift to private or internal channels.
• The paper recommends measuring public knowledge contributions through staggered AI rollouts and strengthening open repositories, professional communities, standards bodies, and universities rather than attempting infeasible global limits on AI precision.

Overview

This paper by Kallel and El Louadi offers a critical appraisal of the knowledge-collapse model of Acemoglu, Kong and Ozdaglar (NBER Working Paper 34910), which argues that agentic AI can trigger a self-reinforcing deterioration of humanity's collective knowledge stock. The authors' contribution is twofold: they correct both the model's assumptions and its viral misinterpretation, and they reframe the underlying mechanism as a "tragedy of the cognitive commons" with a constructive measurement and institutional agenda. Their central thesis is that the model's value lies not in predicting guaranteed catastrophe but in identifying a credible negative externality whose remedy is better aggregation of human knowledge rather than deliberate degradation of AI precision.

The structure of the knowledge-collapse mechanism

The original model rests on a dichotomy between general knowledge — a slow-accumulating social stock such as clinical reasoning principles or legal doctrine — and context-specific knowledge, which is local and private. These two forms are assumed complementary: a precise context-specific signal is only useful to an agent who retains the conceptual scaffolding to interpret it. Human cognitive effort jointly produces both kinds of knowledge through an economy of scope: a physician working through a difficult diagnosis generates a private signal about the patient and, as a by-product, a thin public signal that replenishes the collective stock. This joint production creates a learning externality — the individual bears the full cost of effort but captures only the private return.

Agentic AI substitutes for the private signal but contributes nothing to the public one. Higher machine accuracy therefore lowers the marginal private return to human effort, reducing the flow of public signals and eroding the general-knowledge stock $G$, which in turn degrades the reliability of the AI recommendations that depend on that same substrate. Crucially, collapse is conditional: it occurs only when human effort is sufficiently elastic with respect to AI accuracy ($\varepsilon$ large). If professional norms, licensure, or curiosity sustain effort regardless of machine performance, the catastrophic equilibrium is not reached.

The model's most counterintuitive result is the non-monotonicity of welfare in AI accuracy. Welfare rises while the marginal static surplus from personalized recommendations exceeds the marginal dynamic loss from a degraded commons, then falls once the inequality reverses. The correct reading is an interior socially optimal accuracy $a_{\text{opt}}$ strictly below the technological maximum — not the popular slogan that "more precision is always worse," which corresponds only to the implausible special case in which the dynamic term dominates everywhere. Kallel and El Louadi stress this distinction because a threshold claim is calibratable and falsifiable, whereas monotone immiseration is not.

They also provide a heuristic reduction of the feedback loop, expressing the commons' law of motion as

$G' = (1-\delta)G + \lambda\,\mathbb{E}[e^*(a,G)]$

where $\delta$ is depreciation of general knowledge and $\lambda$ the productivity of the public signal. Under strong complementarity and high elasticity, the steady-state map becomes S-shaped, admitting multiple fixed points; collapse is the loss of stability of the knowledge-rich equilibrium once accuracy crosses a critical value $a^\star$ increasing in $\delta$ and $\varepsilon$ and decreasing in $\lambda$. Three parameters govern everything: effort elasticity $\varepsilon$0, depreciation $\varepsilon$1, and public-signal productivity $\varepsilon$2 — none of which the source paper measures.

Empirical signals: convergent but local

The authors assemble four strands of evidence that lend the mechanism plausibility without establishing societal-scale collapse.

Public-signal thinning: del Rio-Chanona et al. report roughly a 25% decline in Stack Overflow activity in the six months after ChatGPT's release, relative to comparable platforms and restricted-access regions. What disappears is the act of reasoning in public, exactly the externality-generating behavior the model predicts. However, the decline is equally compatible with relocation of reasoning to less visible venues — private chats, internal wikis — meaning erosion of the commons cannot be distinguished from displacement of measurement by any single platform metric.

Cognitive substrate effects: Kosmyna et al.'s EEG study found lower neural connectivity in LLM-assisted essay writers, many of whom could not reproduce essays they had written, a pattern the authors call "cognitive debt" (Kosmyna et al., 10 Jun 2025). Kallel and El Louadi treat this cautiously: it is a small short-horizon laboratory study, the neuroscientific interpretation is disputed, and it evidences at most the demand-side channel rather than proving it.

Model collapse analogy: Shumailov et al.'s demonstration that generative models trained recursively on their own outputs lose distributional tails is suggestive but fundamentally disanalogous — model collapse is a statistical property of recursive training, whereas the human mechanism is driven by motivation and institutions.

Countervailing reallocation evidence: field studies show substantial productivity gains from generative AI, concentrated among less-experienced workers (Brynjolfsson et al.; Noy & Zhang). These measure private productivity, not the public stock, but they reveal a reallocation channel the model omits. The authors elevate this caveat to a potential falsifier: if freed effort systematically flows into public-signal-generating activity, the central mechanism fails rather than merely weakens.

The honest summary they offer is that the mechanism is possible and observable locally, its aggregate sign unknown, and no study observes the full feedback loop closing at societal scale.

Five structural criticisms

A frozen taxonomy of knowledge. The model's most structural weakness is its fixed dichotomy between general and context-specific knowledge. Cognitive technologies historically generate new categories of competence, and a static taxonomy can score only losses, never the emergence of hybrid human–machine capabilities. The authors propose disaggregating $\varepsilon$3 by cognitive level — declarative, procedural, causal, metacognitive — since agentic systems substitute most easily for declarative recall and least for metacognitive judgment. They also note that any taxonomy embeds value judgments about whose expertise counts.

AI can produce general knowledge. The assumption that agentic AI recycles but never produces general knowledge is increasingly strained: AlphaFold's protein-structure predictions are permanent public contributions to biological understanding. The authors extend the law of motion with a machine-contribution term,

$\varepsilon$4

and observe that AI may also raise $\varepsilon$5 itself by lowering the cost of converting private insight into publishable or reusable form. With $\varepsilon$6, the steady-state stock need not fall in accuracy at all. They conclude that the model's most consequential assumption, $\varepsilon$7, is also its most empirically vulnerable.

Effort elasticity is unmeasured. Everything hinges on $\varepsilon$8, which varies across domains with licensure, liability, reputation, and intrinsic motivation. Whether the economy as a whole sits in the collapse region is an empirical question the theory frames but cannot answer.

The long history of cognitive alarm. Socrates on writing, and subsequent panics over printing, telephony, and computing, all proved false positives. The authors concede one genuine disanalogy: agentic AI is the first technology to substitute for the inferential step itself rather than merely storing or transmitting its products (citing Hutchins and Clark & Chalmers). But the base rate of failed alarms should lower the prior on any new instance.

Political futility of throttling precision. No developer will degrade its own product; unilateral restriction is competitively self-defeating; coordinated international caps on model accuracy face formidable technical and verification obstacles, and states with leading capability have little incentive to agree. Sectoral information-design rules are feasible but partial. The second horn of the recommendation — strengthening aggregation institutions — is far more promising.

The authors also turn the critique on themselves, conceding that their evidence base is purposive rather than systematic, that their formal sketch is a heuristic reduction rather than a verified derivation, and that the working paper under review may be revised.

Calibration and falsifiability

The paper insists that a model earns its place in forecasting through testable content. It derives three qualitative predictions: public-signal generation should decline sharply in instrumentally motivated domains but little in professionally bound ones (heterogeneity across domains); the decline should concentrate on the public component of effort rather than private task completion; and, most distinctively, AI recommendation quality should degrade in domains past the threshold, since the thinning commons feeds back into the systems that rely on it. This last prediction is the theory's most valuable because the optimistic co-evolution scenario would never make it.

Falsification would follow from: no decline in public-signal generation after accounting for relocation; systematic redirection of freed effort toward public contributions; AI adding to the general stock at rates matching human contributions; or stable recommendation quality in high-penetration domains.

For measuring $\varepsilon$9, the authors propose exploiting staggered AI-assistant rollouts across firms and jurisdictions with difference-in-differences or event-study designs, using heterogeneity-robust estimators given the known biases of two-way fixed effects (de Chaisemartin & D'Haultfoeuille; Callaway & Sant'Anna). Outcomes must target the public dimension of effort — repository contributions, externalized reasoning, documentation — and separate intensive-margin reduction from extensive-margin exit. Two cautions are registered: adoption-based designs recover at best a reduced-form contribution elasticity, not the structural parameter, requiring exclusion restrictions drawn perhaps from licensing regimes; and short-run event-study elasticities may diverge substantially from the long-run elasticities that drive steady-state dynamics.

Separating the model from its viral image

The popular compression of the paper into headlines — that collapse is guaranteed, that AI will destroy human knowledge, that more precise AI is always worse, that a Nobel laureate has proven doom — misstates the result at every point. Collapse holds only under high elasticity and high precision; the model concerns the collective stock, not individual cognition; welfare is non-monotonic with real gains at low precision; and the Nobel citation honored work on institutions, which if anything points toward institutional remedies rather than technological fatalism. The authors argue that correcting these slippages is a precondition for taking the model seriously, since a precisely stated hypothesis can be tested while a slogan can only be believed.

What survives and the forward agenda

What endures is the diagnosis of a genuine negative externality: when learning is a public good produced as a by-product of private effort, and AI severs the link between private benefit and public contribution, the market under-supplies general knowledge. This places the problem squarely in the tradition of Hardin, Ostrom, Hess and Ostrom, and Romer's knowledge spillovers. The preferred policy response operates on the human side: communities of practice, standards bodies, open repositories, and universities as stewards of validated knowledge raise the productivity of public-signal generation without requiring developers to sabotage their products or states to coordinate internationally.

The research agenda comprises five items: estimating the cross-domain distribution of effort elasticity via quasi-experimental designs; building a dashboard of cognitive-commons health indicators (with persistence, not transient fluctuation, as the criterion), connecting to Woolley et al.'s collective-intelligence factor; modeling AI's endogenous contribution to the general stock as a design problem; studying incentive-based compensation mechanisms such as proof-of-contribution schemes; and focusing governance on feasible, national or sectoral aggregation institutions rather than global precision caps. Two theoretical questions underpin the agenda: the functional form of the complementarity $a_{\text{opt}}$0, which determines whether collapse is gradual or abrupt, and whether aggregation institutions and precision regulation are substitutes or complements.

For forecasting specifically, the methodological lesson is that task-level productivity aggregation is blind to stock effects on knowledge and will be biased optimistic precisely when the dynamic externality binds. The appropriate output is scenario analysis along three branches — cognitive renaissance, stagnation, and domain-limited collapse — with branch probabilities contingent on the elasticity estimates above. The authors also flag a geopolitical asymmetry: general knowledge flows across borders, so a country can free-ride on others' commons, but tacit and context-specific layers do not travel, making dependence on imported general knowledge a sovereignty risk.

Limitations and open questions

Beyond the self-acknowledged limits noted above, several questions remain open. The identification of $a_{\text{opt}}$1 requires strong exclusion restrictions and is acknowledged as the hardest parameter to estimate. Whether the complementarity takes multiplicative, CES, or sharply nonlinear form — and hence whether any decline would be gradual or discontinuous — is unresolved. Whether relocation of reasoning off observable platforms constitutes erosion or mere measurement displacement cannot be settled with existing data. And whether aggregation institutions and any form of precision governance interact as complements remains formally unexamined.

Conclusion

Kallel and El Louadi position the knowledge-collapse model as neither proof of imminent disaster nor alarmist overreach, but as a conditionally stated hypothesis resting on an unmeasured elasticity and two increasingly strained assumptions. Its durable contribution is the identification of a learning externality that conventional productivity-focused forecasts structurally omit, together with a policy prescription — invest in institutions that aggregate and validate human knowledge — that is robust across parameter values. Whether the cognitive commons thins or thickens is, on their account, largely a matter of institutional choice, which is why the model merits calibrated scrutiny rather than either fatalism or dismissal.