AI Can Drain the Expert Pipeline Before Anyone Notices

AI can raise today’s output while quietly removing the junior work that creates tomorrow’s experts.

Source note: Nolan Lovett. “The Tragedy of the Cognitive Commons: How AI Could Disrupt the Regeneration of Professional Expertise.” Human Resource Development Review, published online July 26, 2026. Author accepted manuscript, submitted to arXiv July 31, 2026.

Why This Paper Matters

The easiest way to measure AI adoption is to count what it produces now: faster analysis, more code, shorter research cycles, fewer hours per task. Nolan Lovett’s paper asks a slower and more uncomfortable question: who will be able to judge that work ten years from now?

Many professions regenerate expertise through junior work. New lawyers learn by researching cases and drafting documents. New analysts learn by building models and investigating companies. New software engineers learn by writing, testing, debugging, and maintaining code. Those tasks are productive work for the employer, but they are also the practice through which novices acquire mental models, pattern recognition, and judgment.

AI can perform or accelerate exactly this kind of work. That creates an attractive bargain for each firm. It can remove junior roles, reduce supervision costs, and still hire experienced people later when difficult judgment is needed. The bargain works for one firm because the labor market contains experts trained by many other firms over the previous decade.

The problem appears when every firm makes the same calculation. Each organization saves money, while the profession loses part of the pipeline that produces its next generation of experts. Lovett calls that shared pool of deep professional expertise the Cognitive Commons.

The paper does not claim that professional expertise has already collapsed. It describes a risk that ordinary productivity and headcount metrics can miss: the stock of senior experts may look healthy for years after the flow of new experts begins to weaken.

The Idea in Plain English

Think of professional expertise as a reservoir.

Experienced practitioners are the water already stored. Entry-level roles, supervised practice, difficult assignments, correction, and repetition are the inflow. Organizations draw from the reservoir whenever they hire a senior worker, ask an expert to validate an AI output, or depend on professional judgment during a crisis.

AI can increase the rate at which organizations draw value from the reservoir while reducing the inflow. Existing experts remain available, so the system looks stable. The damage becomes visible only when experienced workers retire and there are too few capable successors.

Lovett argues that the resource behaves like a commons for three reasons:

1. Organizations depend on it collectively. A profession-wide pool of experts supports hiring, regulation, crisis response, consulting, and knowledge transfer.
2. Its benefits are hard to exclude. A firm that stops training juniors can still recruit experts developed elsewhere.
3. Its regeneration can be degraded. If enough firms remove the work that builds expertise, the shared pool eventually shrinks or becomes shallower.

The paper adds a second distinction. AI-era work requires both Internalized Mastery and Distributed Mastery.

Internalized Mastery is deep knowledge held by a person: the schemas, tacit understanding, and diagnostic judgment built through sustained practice. Distributed Mastery is the ability to orchestrate work across people and AI systems: prompting, comparing, curating, and designing workflows.

Distributed Mastery cannot safely replace Internalized Mastery. Lovett’s Validation Tether says effective AI orchestration still depends on enough internalized expertise to recognize when a polished output is wrong in a way only a domain expert can see.

Evidence Behind the Framework

This paper does not report a new experiment. It is a conceptual synthesis that combines four bodies of work:

• Commons theory, especially the difference between using a shared resource and regenerating it.
• Human resource development and adult-learning research on how expertise forms through workplace practice.
• Distributed cognition research on how capability is shared across people, tools, and systems.
• Systems thinking about stocks, flows, feedback loops, and long delays between cause and consequence.

Lovett uses existing labor-market, clinical, organizational, and experimental studies to establish pieces of the proposed mechanism. The strongest empirical evidence concerns narrower claims: early-career employment declines in some highly AI-exposed occupations, better performance while AI is available without equivalent independent skill transfer, and failures to detect or override flawed AI outputs.

The larger claim, that these patterns will produce profession-level depletion of expertise, remains a structural prediction. The paper explicitly separates what has been observed from what the framework predicts.

It also specifies when the framework should apply. Three conditions must occur together:

• AI substitutes for tasks that historically helped novices develop expertise.
• Those developmental experiences are difficult to reproduce outside real organizational work.
• The occupation depends on a shared, non-exclusive expertise pool whose regeneration can fail.

The occupation is therefore the right unit of analysis.

What the Framework Proposes

The hidden subsidy for training was junior work itself

Organizations have always had weak incentives to fund general professional development because trained workers can leave. Yet firms still developed experts because junior employees were needed to perform junior tasks. Learning happened alongside production.

AI could break that accidental alignment. If a system handles the production task, a firm may no longer need the novice whose work once justified the cost of supervision. The loss goes beyond training budgets because it removes the productive activity through which much training occurred.

Depletion can happen through fewer roles or shallower learning

The paper proposes two mechanisms.

The first is direct position elimination. An organization removes junior roles because AI can perform enough of their work.

The second is augmentation without internalization. Junior workers remain employed and produce more with AI, but the system supplies finished outputs before they struggle with the underlying problem. Performance improves while the worker has access to the tool, yet the mental models needed for independent judgment may not develop to the same degree.

Augmentation without internalization is harder to spot. Headcount and output can look healthy even while the developmental quality of work declines.

Oversight has a shallow layer and a deep layer

Lovett distinguishes surface validation from substantive validation.

Surface validation catches bad formatting, broken logic, obvious contradictions, and implausible answers. People can learn much of it through experience with AI tools.

Substantive validation catches domain-specific mistakes inside otherwise coherent work: a risk model built on the wrong assumption, code that runs but creates architectural debt, or a diagnosis that fits a pattern while ignoring a patient-specific fact. This requires independent domain knowledge.

A human in the loop can still fail as a safety mechanism. A person can be present, comfortable with the interface, and capable of surface review while lacking the expertise or epistemic confidence to challenge the system at the point that matters.

The warning signs are real, but early and uneven

The paper cites Stanford Digital Economy Lab payroll research that found a 16 percent relative employment decline from October 2022 to September 2025 among workers aged 22 to 25 in the most AI-exposed occupations, while employment for workers aged 35 to 49 in those occupations grew by more than 8 percent. Lovett treats this as an early signal consistent with pipeline disruption, not proof of a profession-wide collapse.

Other studies reviewed in the paper report reduced independent performance after AI-assisted work, limited transfer from assisted to unassisted tasks, and cases in which people noticed errors but still followed flawed AI advice. These findings support parts of the mechanism. They do not yet establish the long-term commons outcome.

Vulnerability differs sharply by occupation

The framework proposes five factors that shape risk:

• Task substitutability: whether AI can do the work novices historically learned from.
• Regulatory intensity: whether supervised practice and demonstrated human competence are institutionally required.
• Safety criticality: whether failures are visible and consequential enough to preserve accountability.
• Professional-association strength: whether a profession can coordinate standards and developmental obligations.
• Work modularization: whether work can be broken into discrete tasks that AI can absorb without integrated human judgment.

The paper therefore expects greater vulnerability in occupations such as software engineering, financial analysis, and legal research than in fields with stronger licensing, supervised-practice requirements, or embodied learning.

Why It Happens

The coordination failure unfolds slowly.

Each firm captures nearly all of the immediate benefit from reducing junior labor: lower salaries, lower training costs, and faster production. The cost of producing fewer future experts is spread across the entire profession. No single hiring decision changes the market enough to justify one firm carrying the training burden for everyone else.

The current supply of senior workers also disguises the problem. Experts available in 2026 were developed by investments made years earlier. A firm can eliminate entry-level roles today and still recruit experienced workers tomorrow. If many firms do this, the shortage may not become obvious until the 2030s or 2040s.

The paper calls part of this the Human Reserve Paradox. Organizations want experienced people available for exceptions, crises, and AI validation, but each has an incentive to avoid paying to maintain that reserve. The value of the reserve is most visible only after something fails.

Validation becomes fragile for the same reason. As AI use grows, organizations need more people who can distinguish credible output from correct output. But if AI removes the practice that builds that judgment, adoption weakens its own safety layer.

What This Means for Builders

AI products shape expertise development even when their teams do not think of themselves as educators.

A system that immediately generates a finished answer optimizes for present output. A system that requires an initial human attempt, asks the user to compare alternatives, exposes uncertainty, or turns explanations into questions can preserve more of the cognitive work that supports learning.

Lovett does not name a product pattern called “developmental mode.” That is a practical extension of the paper’s argument: builders could offer a learning-oriented path alongside the fastest production mode. Useful design questions include:

• Does the workflow ask the practitioner to form a view before seeing the model’s answer?
• Can the system reveal evidence and uncertainty without resolving every judgment automatically?
• Does review require an explicit rationale rather than a passive approval click?
• Are users periodically tested on comparable work without assistance?
• Can product telemetry separate corrections of superficial defects from detection of substantive domain errors?

These patterns will not recreate apprenticeship by themselves. They can still preserve some of the struggle through which judgment develops.

Builders also need to resist treating human approval rates as proof of reliable oversight. If reviewers share the same blind spots, or if their expertise has never developed independently, a high approval rate can measure deference rather than quality.

What This Means for Buyers and Operators

The practical risk is optimizing the workflow while neglecting the pipeline that makes the workflow trustworthy.

Buyers should evaluate AI programs on two time horizons. The first is immediate productivity. The second is whether the operating model still produces people capable of independent performance and substantive validation.

Leaders should therefore ask:

• Which tasks in this workflow are merely low-value, and which are formative for future experts?
• What work will a junior employee do to acquire the judgment currently held by senior staff?
• Does AI assistance arrive before or after the employee attempts the problem?
• Can practitioners still perform representative tasks without AI?
• Who owns the cost of maintaining the profession’s developmental pipeline?

Lovett proposes phased AI access, protected practice without AI, minimum periods of independent work, active comparison of human and model reasoning, and assessment that separates assisted output from unaided capability. The specific interface patterns above are our translation of those principles into product design, not labels used in the paper.

Professional associations or policy may need to address the profession-wide coordination problem. Associations can maintain supervised-practice standards, monitor changes in entry pathways, and experiment with credentials that distinguish AI-assisted competence from demonstrated Internalized Mastery. Public support could make developmental roles less costly in occupations where the market would otherwise underprovide them.

The goal is to preserve expertise regeneration while job structures change, instead of assuming it will continue as a byproduct of work.

What to Watch Next

Longitudinal evidence will matter most. Researchers need to follow cohorts from entry level into mid-career and compare their independent performance, validation ability, and developmental experiences across AI-heavy and AI-light workplaces.

Researchers should also track divergence between assisted productivity and independent judgment. If output rises while unaided performance or error detection falls, the Validation Tether is weakening even before headcount shortages appear.

Comparisons across occupations and countries will test the framework’s boundary conditions. Strong apprenticeship systems, licensing regimes, and professional associations may preserve regeneration better than market-mediated pathways. AI-supported mentorship or new credentialing systems may also create viable replacement pathways that do not resemble traditional entry-level jobs.

Professional bodies also need better commons-health indicators. Entry-level hiring ratios, independent performance assessments, seeded-error detection tests, training investment, and the age distribution of validation-capable experts could reveal pipeline damage earlier than senior hiring shortages.

Limitations and Caveats

The framework integrates evidence and makes falsifiable predictions, but it does not directly observe the profession-level depletion it warns about.

The labor-market evidence covers only the first few years of widespread generative AI adoption. The strongest employment signals come from highly exposed occupations and should not be generalized to all professions. Some large studies find mixed, adaptive, or null labor-market effects over the same early period.

The framework also assumes that consequential AI systems will continue to require human oversight. If systems become reliably autonomous across a domain, the problem changes from expertise regeneration to workforce displacement. That is a different argument.

Traditional entry-level employment may not be the only way to build expertise. New apprenticeships, AI-mediated coaching, simulation, and different forms of credentialing could replace some lost learning opportunities. The paper identifies this as an open empirical question rather than ruling it out.

The Western professional model also shapes much of the evidence. Guilds, community-based learning, and national apprenticeship systems may face different dynamics.

Some AI workflows preserve productive struggle by requiring an independent attempt, active evaluation, or justification. The test is how much cognitive work the system leaves for the person.

Source

Lovett, N. (2026). The Tragedy of the Cognitive Commons: How AI Could Disrupt the Regeneration of Professional Expertise. Human Resource Development Review, advance online publication. Author accepted manuscript available as arXiv:2607.29380. https://arxiv.org/abs/2607.29380. Version of record: https://doi.org/10.1177/15344843261470602