The Tragedy of the Cognitive Commons. How AI Disrupts the Regeneration of Expertise

Every company can make the same reasonable decision. Automate the junior work now. Hire experienced people later, when judgment is needed.

The problem appears when every company follows that plan.

Experienced people are not a natural resource waiting in the market. They became experienced because someone once paid them to do slower work, make mistakes, receive correction, and take on harder problems. Remove that path across a profession and the market may look healthy for years while its ability to produce the next generation is already failing.

Nolan Lovett calls the shared pool of deep professional expertise the cognitive commons. His argument is uncomfortable because organizations still depend on this commons precisely when AI looks most capable.

The experts in the market came from somewhere

A cognitive commons is the distributed body of people who have internalized a field deeply enough to recognize when a plausible answer is wrong in context.

No company owns this pool. A firm can hire from it, consult it, and rely on the standards it maintains. The benefits are shared while the cost of training a beginner is paid by particular organizations and by the beginner.

That creates a familiar collective action problem. A company that removes entry-level roles receives an immediate efficiency gain. It can still recruit experts trained elsewhere. The loss appears across the profession and arrives with a long delay.

Today's senior engineer, doctor, lawyer, or analyst carries development that began years ago. A decline in junior opportunities can remain hidden until the current cohort ages out and the replacement cohort turns out to be smaller or shallower.

Surface validation is not enough

The paper makes a useful distinction between surface validation and substantive validation.

Surface validation checks whether an output is coherent, properly formatted, and internally consistent. A person can become good at this through repeated interaction with AI. The model says something odd, the user spots the visible flaw, and the prompt improves.

Substantive validation asks whether a mathematically correct risk model rests on the wrong market assumption, whether a plausible diagnosis ignores a patient-specific factor, or whether working code creates an architectural failure six months later.

Those errors do not announce themselves. Detecting them requires knowledge that has become internal. The validator needs an independent model of the field, not only fluency in operating the AI system.

Lovett calls this dependency the validation tether. Distributed mastery, which includes prompting, orchestration, and human-AI workflow design, remains tied to internalized mastery when the output matters.

The two forms of expertise can exist in the same person. That combination may become one of the most valuable profiles in AI-intensive work. The danger is assuming that the first can permanently replace the second.

Efficiency can hide degradation

The paper describes two ways the commons can weaken.

The visible one is stock depletion. Fewer beginners enter, so the profession eventually has fewer people with deep expertise.

The quieter one is functional degradation. Junior workers remain employed and produce at a high level with AI, but they skip part of the cognitive struggle through which earlier cohorts learned the field. Their output looks senior before their judgment becomes senior.

This second failure can coexist with strong productivity numbers. Surface checks continue to pass. Organizations become more confident in AI at the same time that fewer people can recognize its subtle failures.

The paper is conceptual and its strongest long-term claims still need direct testing. It also cites counterevidence showing that AI workflows can improve expertise when they require active human participation. That qualification changes the practical conclusion.

The question is not whether people use AI. It is how the work is designed around it.

Preserve the struggle that builds a model

Some parts of learning should remain temporarily unaided. A junior worker may need to form an initial diagnosis before seeing the model's answer. A review may need to ask for reasoning, not only accept a polished result. A team may need protected practice where speed is not the only measure.

This does not require preserving every inefficient ritual. It requires knowing which friction is waste and which friction is education.

The same distinction applies personally. Learning to orchestrate agents is useful. It does not automatically create the domain judgment needed to verify them. If I become better at directing AI while allowing my underlying knowledge to decay, the system can make me look more capable and leave me less able to know when it is wrong.

AI can expand access to professional performance. Whether it also expands professional expertise depends on the training paths, review structures, and protected struggle we choose to keep.

The cognitive commons will not replenish itself simply because the outputs still look good.

Source paper by Nolan Lovett, The tragedy of the cognitive commons.