The Tragedy of the Cognitive Commons: Why Expertise Is a Shared Resource, and Why It Can Be Overgrazed

In 1968 the ecologist Garrett Hardin described a pasture open to every herder in a village. Each herder benefits individually from adding one more animal to the commons, while the cost of overgrazing is spread across everyone. Acting rationally as individuals, the herders collectively ruin the resource they all depend on. The essay was written about grazing land, fisheries, and clean air, but the underlying structure applies to something less tangible and, arguably, more consequential: the pool of deep, hard won expertise that professions and societies draw on to function. Economists and human resource researchers have recently started calling this pool the cognitive commons, and the emerging concern, sometimes framed as the tragedy of the cognitive commons, is that the mass adoption of generative artificial intelligence is quietly grazing it down.

What the cognitive commons actually is

The cognitive commons is not a metaphor for information. Information is abundant, searchable, and largely non rivalrous, you can copy a fact without diminishing anyone else's access to it. The cognitive commons refers to something narrower and more fragile: the collective capacity to generate, judge, correct, and pass on sound reasoning within a domain. It lives in the intuitions of a surgeon who has made a thousand small judgment calls in the operating theatre, in the pattern recognition of a structural engineer who has seen a beam fail, in the editorial instincts of a journalist who knows which source is lying before the fact check confirms it. None of that is written down in a form that fully captures it. It is renewed the way any commons is renewed, through use, apprenticeship, and the slow correction of error by people who already have the underlying competence to notice when something is wrong.

This is why the commons framing is more accurate than simply calling expertise a resource. A resource can sit in storage. The cognitive commons cannot. It has to be continuously regrazed and regrown by the very people who rely on it, or it degrades. A textbook of settled procedures is not the commons, it is the fence around what has already been harvested from it. The commons is the ongoing judgment that produces new settled procedures when the old ones stop working.

The mechanism of depletion

The economic argument for why artificial intelligence adoption threatens this commons is not that the tools produce bad answers. Often they produce serviceable ones. The threat is structural, and it mirrors Hardin's original logic almost exactly. Any individual professional, student, or organization has a strong short term incentive to delegate a reasoning task to a model rather than work through it themselves, because the delegation is faster and the individual cost of skipping the underlying practice is invisible in the moment. But the practice itself, the years of making small judgment calls and getting some of them wrong in low stakes settings, is precisely how expertise is built and renewed in the first place. When enough individuals make the locally rational choice to delegate, the collective supply of people who can actually verify, extend, or correct the output of these systems starts to shrink. Nobody intends to deplete the commons. Everybody is simply grazing their own animal.

Erik Dane's research on cognitive entrenchment offers an important companion insight here. Deep expertise is not purely additive, more experience does not always mean more flexibility. Highly entrenched experts can become rigid within their own domain even without any AI involvement. What generative tools change is the entry pipeline beneath that expertise, the population of newer practitioners who would otherwise be accumulating the very experience that later lets them catch their own entrenchment and correct for it. If fewer people pass through the effortful, error prone stage of skill formation, there are fewer people positioned to notice when a mature expert, human or automated, has quietly become wrong.

Why this is difficult to see coming

Commons depletion is notoriously hard to detect early because averages hide it. A profession can look perfectly healthy in the aggregate output it produces this year, this quarter, even this decade, while the underlying reserve of people capable of independently generating that output without assistance is thinning. The parallel to overfishing is instructive. Catch volumes in a depleted fishery often stay stable or even rise for a period before collapsing, because the industry compensates with more boats and better sonar, extracting the same yield from a smaller breeding population until the population can no longer replace itself. Applied to expertise, the equivalent compensation is more tool assisted output per practitioner, which conceals the fact that the practitioner base capable of overseeing that output without the tool is contracting.

This connects directly to what has been called the HAL dilemma in AI safety discourse: the more obedient and capable a system becomes, the more dangerous it is to place it beyond meaningful human oversight, because obedience without context or judgment can execute a flawed instruction with perfect precision. The cognitive commons is the thing that supplies that oversight capacity. If the commons thins at the same rate that reliance on capable systems grows, the safety margin narrows from both directions simultaneously. This is a different and more structural problem than the usual worry about AI making mistakes. It is a worry about who, in ten or twenty years, will still have the trained judgment to recognize the mistake at all.

A historical parallel worth taking seriously

Guild systems in early modern Europe understood something about this problem that is easy to lose sight of now that credentialing has become bureaucratic rather than relational. The Amsterdam Surgeons' Guild, for instance, commissioned group portraits of its senior members conducting public anatomy and osteology lessons, not merely as vanity pieces but as records of a specific pedagogical event, a public demonstration in which senior expertise was transmitted in front of witnesses whose own reputations were staked on having been present and having understood it. The apprentice did not receive a manual. He received supervised exposure to a living expert working through an actual case, with the freedom to ask questions and the obligation to eventually be tested by peers who had been through the same process. The guild's entire authority rested on the credible claim that its members had personally undergone this transmission chain, not that they had access to accumulated guild knowledge in the abstract.

The alchemical workshop tradition offers a related, messier case. Much of early modern alchemy was genuinely unreliable science mixed with real proto chemical technique, and the only way apprentices learned to tell the difference, which furnace temperature was safe, which reaction was worth repeating, which "success" was actually contamination, was by standing in the workshop next to someone who had already made the relevant mistakes. There was no compressed shortcut to that judgment. Reading a treatise on distillation was not the same as having watched a real vessel crack.

Neither of these historical models scales cleanly to the present, and nobody should be nostalgic for guild gatekeeping or alchemical trial and error as institutions. But they illustrate a point that economic research on the cognitive commons is now making with more rigorous language, expertise has always depended on structured, witnessed, effortful transmission between a person who has it and a person who does not yet have it. Compressing or automating that transmission step does not preserve the expertise, it preserves only its current output.

Where this leaves practice, not just theory

None of this is an argument against using capable tools. It is an argument for treating the effortful, occasionally frustrating parts of skill formation, the parts most tempting to delegate, as a resource worth protecting rather than a cost worth eliminating. Grandomastery approaches a version of this problem from the language and creativity side rather than the professional expertise side, but the underlying commitment is the same, the platform's activities are handcrafted and deliberately non AI generated, precisely because the founder, Alexander Popov, has argued that outsourcing the moment of struggle removes the mechanism by which fluency and judgment actually develop, even when the delegated output looks fine on the surface. Activities such as Random Profession, which asks a learner to justify a drastic and implausible career shift using only their own reasoning, or Random Case File, which puts a learner in the position of building and defending an argument under adversarial questioning, are built around exactly the kind of unassisted, witnessed reasoning under uncertainty that expertise transmission has always required, just relocated to language and abstract thinking rather than to a surgical amphitheatre or a guild workshop. The main platform frames this explicitly as training for a world where pattern interpolation is cheap and independent judgment is not.

The tragedy in tragedy of the commons is not inevitability, Hardin's own later work acknowledged that commons can be managed successfully through deliberate coordination rather than privatization or collapse, as Elinor Ostrom demonstrated empirically across dozens of real world cases of shared resource governance. The cognitive commons is no different in kind. It requires people and institutions to notice that convenience has a cost that does not show up on this quarter's ledger, and to keep investing effortful, unassisted practice into the pool even when a faster substitute is sitting right there. The alternative is a professional world that looks fully staffed and fully capable for a long time, right up until the moment it discovers that the people who could have caught the error are no longer in the room.