A new argument about AI and work focuses less on immediate job losses and more on what happens years later. Nolan Lovett of the NATO Special Operations University describes a risk he calls the "tragedy of the cognitive commons": individual companies can make rational decisions to use AI, while the professional talent pool they all depend on becomes weaker over time.
The concern is not that AI is always harmful. It is that some forms of AI adoption may remove the learning path that turns beginners into experts. If enough organizations make the same choice, the result could be a profession with fewer people able to judge, correct, or override AI systems when it matters.
Why entry-level work matters
Lovett builds the idea from the tragedy of the commons, the dilemma described in 1968 by ecologist Garrett Hardin. In that model, each herder benefits from adding one more animal to a shared pasture, while the damage to the pasture is spread across the group. Each individual decision is rational, but the shared resource is degraded.
In Lovett's version, the shared resource is professional expertise. When a company replaces entry-level positions with AI, it receives 100 percent of the efficiency gains. The cost, however, lands across the wider labor market because all organizations draw from the same pool of trained people.
That matters because experienced professionals do not appear instantly. They are formed through years of lower-level work, repeated judgment calls, errors, corrections, and gradually harder assignments. If AI absorbs the tasks that once trained junior workers, the system that renews expertise can start to fail.
The hidden cost of AI assistance
Lovett identifies two main ways the pipeline can break. The most direct is the elimination of entry-level positions, where AI systems take on work that would otherwise be handled by junior employees.
The second route is more subtle. Entry-level jobs may remain, but AI assistance can let junior workers reach productivity levels that previously required years of experience. That can look like success in the short term. But if the worker is no longer doing the hard cognitive labor underneath the result, deep domain knowledge may not develop in the same way.
This creates a difficult management problem. AI can make work faster while also reducing the practice that teaches people how to do that work independently. The immediate metric improves, while the long-term capability becomes harder to see.
The validation tether problem
Lovett calls the next issue the "validation tether." AI systems still need human oversight, especially when their outputs look convincing but contain domain-specific mistakes. The people best equipped to catch those errors are the same experts whose development may be weakened by overreliance on AI.
Surface-level review is not enough for this kind of supervision. According to the research Lovett discusses, detecting serious problems in plausible AI output requires deep knowledge of the field, not just the ability to spot obvious contradictions.
There is also a habit problem. People who often treat AI answers as dependable can lose the reflex to challenge them. At the same time, the older training settings where junior workers learned to question authority and test claims against reality may shrink as entry-level work changes.
Why the damage may arrive late
The risk is delayed because the current labor market still contains people trained before the newest wave of AI adoption. Lovett argues that today's experienced professionals were trained 5 to 20 years ago. If entry-level positions began being cut starting in 2023, the full effect may not appear until somewhere between 2030 and 2045.
He calls this the "Human Reserve Paradox." Organizations need expert humans in reserve for validation, crisis management, and cases where AI systems are overwhelmed. Yet no single organization has enough economic incentive to maintain that reserve alone.
Even workers who remain in the pipeline could arrive with thinner expertise, Lovett argues, if their careers are built around orchestrating AI rather than performing independent cognitive work. The profession may still have workers, but fewer of them may have the depth needed when automation fails or uncertainty rises.
Which fields face the most pressure
Lovett does not claim every profession faces the same level of exposure. He places software engineering, financial analysis, and legal research in the highest vulnerability category because they combine high task substitutability, relatively light regulation, and strong modularity.
Medicine and engineering have some protection from stricter regulatory requirements and stronger professional associations, but they are not immune. The difference is not whether AI can be used, but whether the profession has structures that preserve training, standards, and independent competence.
Lovett's proposed responses do not center on banning AI. Instead, he argues for AI-free learning environments, phased AI introduction, and a baseline of human performance before AI is added. He also recommends that professional associations test domain competence through certifications alongside AI skills, and that policymakers make training and education more attractive.
What the evidence shows so far
The labor-market evidence is still mixed. A study from summer 2025 cited by Lovett showed employment declines in AI-affected occupations, especially among young workers. For more experienced workers in those fields, employment held steady or even grew, with researchers pointing to AI replacing codified knowledge while practical experience remained valuable.
A Federal Reserve Board study found that growth in programming jobs has nearly halved since ChatGPT launched, though a clear causal link to AI has not been proven. Other forces, including tighter monetary policy and a correction after pandemic-era tech overhiring, may also matter. A January 2026 study found that the job crisis in AI-affected occupations began before ChatGPT's release.
An Anthropic study from March 2026 found no measurable overall labor-market impact from AI. But it did identify one warning sign: among young workers aged 22 to 25 in highly AI-exposed occupations, the job-finding rate dropped by half a percentage point.
The cognitive evidence is clearer in the source material. An MIT study using EEG measurements found that even brief AI use weakened neural connectivity, and over 80 percent of participants struggled to recall content from their own AI-assisted writing. An Anthropic study with software developers found that participants with AI access scored 17 percent worse on knowledge tests, with the largest losses among those who used AI as an answer machine.
A Swiss study of 666 participants found a strong negative link between AI use and critical thinking, most pronounced among 17- to 25-year-olds. Among students in China, homework grades improved by 18 percent, but exam performance dropped by up to 24 percent, with the full effect appearing after about two years.
The practical lesson is narrow but important: the issue is not only how much people use AI, but how they use it. Treating AI as a substitute for thinking appears to carry the greatest risk, while using it for explanations can reduce the negative effects.