The Tragedy of the Cognitive Commons: How AI Could Disrupt the Regeneration of Professional Expertise
Nolan Lovett
cs.CY, econ.GN
2026-07-31
Each firm rationally adopting AI can drain a profession's shared expertise like overgrazing empties a commons; the deep knowledge needed to catch AI's mistakes is exactly what erodes.
Human Resource Development (HRD) scholarship has framed AI's impact on work as an organizational training problem: should employees learn prompt engineering, should companies run reskilling programs. This 2026 paper in Human Resource Development Review argues that framing misses a collective-level question: how professional expertise actually regenerates from one generation of practitioners to the next. Author Nolan Lovett borrows the structure of Hardin's 1968 "tragedy of the commons" and names a "Cognitive Commons", the pool of deep expertise a profession accumulates, shared by all but owned by none, whose supply lines AI adoption may be cutting.
He borrows the structure, not the fatalism. Hardin claimed the commons was doomed; Ostrom later showed empirically that communities sustain shared resources through governance such as boundaries, monitoring, and graduated sanctions. The paper does not claim the tragedy is inevitable, only that current arrangements leave a collective-action problem unaddressed.
This is a theoretical synthesis, not an experiment. It stitches together four literatures that had stayed separate: commons theory (Hardin, Ostrom), HRD research, distributed cognition (Hutchins), and systems thinking. The core is three constructs.
The key split is surface validation (spotting formatting errors and obvious nonsense) versus substantive validation (recognizing errors that look plausible but are wrong in domain-specific ways). The first you can develop just by using AI; the second requires Internalized Mastery. Vicente and Matute's 2023 experiment makes the point sharply: 80.7% of participants detected the AI's bias, yet still followed the biased advice. Detecting a problem is not the same as overriding it.
The paper also lists five factors that determine a profession's vulnerability (task substitutability, regulatory intensity, safety criticality, professional-association strength, work modularization) and a six-node causal chain.
The labor-market evidence has specific numbers. Brynjolfsson et al. (2025), using payroll data covering 25 million U.S. workers (3.5 to 5 million per month), found that in the most AI-exposed occupations, workers aged 22 to 25 saw a 16% relative employment decline between October 2022 and September 2025 (controlling for firm-level shocks), while workers aged 35 to 49 grew more than 8%. Hampole et al. reinforced the cohort pattern with 58 million LinkedIn profiles.
| Evidence | Source | Number |
| Systematic cuts to entry-level jobs | Brynjolfsson et al. 2025 | ages 22-25 -16%, 35-49 +8% |
| Doctors' independent skill drops after AI | Budzyń et al. 2025 (Lancet) | endoscopist deskilling |
| AI-assisted training does not transfer | Wiles et al. 2024 | no significant advantage unaided |
| Spotting an AI error, following it anyway | Vicente & Matute 2023 | 80.7% detected the bias |
| Workers stop checking AI | Niederhoffer / Benzing 2025 | 40% got flawed content monthly; 60% don't routinely check |
The author stresses these are early signals, not a collapse already underway. A Danish study (Humlum & Vestergaard 2025) found no significant earnings or hours effect in the first two years, and clinician workflows that required active engagement with AI reasoning actually improved diagnostic accuracy (Everett et al. 2025). The mechanism is conditional, not universal.
The most direct line for AI practitioners: the validation tether means "able to use AI" and "able to tell when AI is wrong" are two different skills, and the second is quietly being displaced by the first. As an organization fills with people fluent in prompting but short on the domain knowledge to catch plausible-but-wrong outputs, more AI adoption raises rather than lowers systemic risk, while everything looks fine on the surface.
The time delay is what makes this hard to see. Entry-level positions cut in 2023 will not show up as an experienced-expert shortage until 2030-2045, when those cohorts should have matured. The labor market looks healthy today because it is still drawing on developmental investments made between 2003 and 2020.
The author is unusually explicit about limits, which is a strength. The data spans under three years, which is early stage rather than mature dynamics. The patterns concentrate in the most AI-exposed occupations and do not characterize professional employment broadly. The paper does not demonstrate widespread validation failures or profession-wide expertise collapse. The profession-level depletion is a structural prediction, not an observed outcome; the author offers it as a falsifiable account of where current incentives lead.
Two more caveats on a careful read. The five-factor framework is qualitative, with no weights or thresholds, so it predicts vulnerability only roughly. And it assumes a Western entry-level-then-promotion model of professional formation; apprenticeships, guilds, and community-based transmission might show very different commons dynamics, which the author acknowledges but does not develop.