A peer-reviewed paper published online on July 26, 2026 argues that the same AI adoption decisions currently saving advertising agencies and marketing teams money may be quietly eliminating the entry-level work through which junior staff historically became senior experts capable of catching AI's mistakes.

The paper, titled "The Tragedy of the Cognitive Commons: How AI Could Disrupt the Regeneration of Professional Expertise," was written by Nolan Lovett of NATO Special Operations University and appeared in Human Resource Development Review, a peer-reviewed journal. According to the author accepted manuscript, the work was submitted on November 2, 2025, revised twice, accepted on July 6, 2026, and published online on July 26, 2026. A separate version was posted to the arXiv preprint server on July 31, 2026. Because Lovett prepared the paper as part of official duties for the United States government, it carries no copyright protection under US law and is freely available.

Lovett's argument does not target advertising specifically. It targets any profession where artificial intelligence performs the routine cognitive tasks that novices once used to build expertise. But the mechanism it describes maps with unusual precision onto trends PPC Land has already documented across paid media, agency operations, and search marketing over the past year: entry-level hiring in AI-exposed occupations falling toward vanishing levels, senior confidence in AI outputs rising faster than the habit of checking them, and a widening gap between what AI tools can produce and what the people deploying them actually understand.

What the paper argues

The paper borrows its title and structure from Garrett Hardin's 1968 essay "The Tragedy of the Commons," which described how individually rational people deplete a shared resource that everyone depends on. Lovett applies that logic to what he calls the Cognitive Commons: the collective pool of deep, validated expertise inside a profession, spread across many practitioners at many organizations, that firms draw on even when they do not directly employ the experts who hold it.

Two conditions make expertise behave like a commons, according to the paper. First, it is collectively depended upon: a company hiring an experienced media buyer, financial analyst, or engineer is drawing on training investments some other employer made years earlier, without having borne that cost itself. Second, it degrades through overexploitation of the very mechanism that renews it. An organization that eliminates entry-level positions and relies on artificial intelligence to fill the gap captures the efficiency gain immediately, while the cost, a shrinking supply of experienced practitioners a decade from now, falls on the whole profession, including on that same organization when it eventually needs to hire someone able to catch an error a model cannot see.

Lovett distinguishes two forms of professional skill. Internalized Mastery is deep domain knowledge built through sustained, difficult practice: the kind of judgment a physician, lawyer, or media strategist develops by struggling with cases over years. Distributed Mastery is fluency in directing AI systems, prompting them, reviewing their output, and stitching their work into a finished product. The paper does not dismiss the second skill. It argues instead that the two are not substitutes, and that the newer one silently depends on the older one for something specific: catching AI's mistakes.

That dependency is what the paper calls the Validation Tether. Surface validation, checking that an AI output looks coherent, is formatted correctly, and does not contradict itself, can be learned quickly and does not require deep expertise. Substantive validation, recognizing that a technically correct recommendation is wrong for this client, this market, or this legal jurisdiction, requires exactly the internalized domain knowledge that entry-level AI adoption tends to remove from the pipeline. Lovett writes that effective oversight of AI "fundamentally depends on the internalized domain knowledge, robust mental models, and tacit understanding" that traditional, slower developmental pathways build, and that organizations eliminating those pathways are eroding the foundation their own AI oversight depends on.

The evidence cited

The paper's strongest empirical anchor is labor market research from the Stanford Digital Economy Lab. Analyzing payroll data covering more than 25 million United States workers, researchers Erik Brynjolfsson, Bharat Chandar, and Ruyu Chen found that in the most AI-exposed occupations, employment for workers aged 22 to 25 fell 16 percent between October 2022 and September 2025, even after controlling for firm-level shocks, while employment for workers aged 35 to 49 in those same occupations grew more than 8 percent over the same period. Occupations with lower AI exposure showed no such age gap. A separate analysis of 58 million LinkedIn profiles, led by Menaka Hampole and colleagues at the National Bureau of Economic Research, reached a similar conclusion using different data: the pattern looks like systematic removal of developmental roles rather than a temporary downturn.

Clinical and experimental research adds a second thread. A multicenter study published in The Lancet Gastroenterology and Hepatology, led by Kasper Budzyń, found that endoscopists who had adopted AI-assisted polyp detection showed reduced independent detection accuracy compared with their own prior baseline once the AI assistance was taken away. Separately, a 2023 study by Lucía Vicente and Helena Matute found that most participants, 80.7 percent, could correctly identify errors in AI-generated recommendations when asked directly, yet continued to follow the flawed advice anyway. The gap between noticing a problem and acting on that judgment, the paper argues, is exactly the gap Distributed Mastery cannot close on its own.

A third data point speaks directly to workplace behavior already familiar to marketing operations. Survey research from Microsoft, cited in the paper, found that 60 percent of employees feel confident enough in AI output that they do not routinely check its accuracy, while separate Harvard Business Review-published research led by Kate Niederhoffer found that 40 percent of full-time employees received substantively flawed AI-generated material in the preceding month, spending close to two hours on average fixing each instance.

Why the mechanism resembles market behavior already documented

The paper's central claim, that eliminating entry-level roles looks rational to any single employer while degrading the wider profession, matches a pattern PPC Land has tracked across agency and platform data through 2026. IAB Australia's 2026 Digital Advertising and Ad Tech Industry Talent Review, published June 2, 2026, found entry-level vacancies had collapsed to just 1 percent of the available job pool in Australia's advertising and ad tech sector, with pressure concentrated in the process-driven roles that historically served as training ground for the profession. According to IAB Australia chief executive Gai Le Roy, continued investment in people and talent pipelines "will be critical" if the industry wants to sustain a strong market over time.

Agency-side survey data tells a related story. Basis, a Chicago-based advertising operations software company, published its 2026 Advertising Agency Report on April 20, 2026, surveying 213 agency professionals. It found that 87.3 percent believed the traditional agency model was already broken or would be within three to five years, with 39.9 percent of agencies reporting layoffs in the preceding twelve months. Confidence in the model's future had fallen below 50 percent for the first time in the survey's history. Agency leaders named AI their top investment priority for a second consecutive year, according to the same report, with 77.7 percent of vice presidents and above planning to increase AI spending even as headcount contracted.

This is precisely the tension Lovett's paper formalizes: the efficiency gain from replacing entry-level labor with AI accrues to the individual agency immediately, while the erosion in the profession's future supply of senior, validation-capable strategists accrues collectively and only becomes visible years later. Whether the outcome favors or undermines total marketing employment remains genuinely contested. PPC Land has separately reported on the Jevons paradox argument that cheaper marketing work, made possible by AI, could expand total demand for marketers rather than shrink it, an argument Box chief executive Aaron Levie made in a December 2025 analysis. Lovett's own paper acknowledges this counter-evidence exists; it cites null earnings effects among Danish generative AI adopters and heterogeneous retrainability findings from National Bureau of Economic Research researcher Benjamin Hyman and colleagues as reasons the depletion mechanism is conditional rather than universal.

A pattern already visible outside advertising

The paper's abstract validation and substantive validation distinction has a real-world illustration PPC Land documented independently. A show cause hearing at Florida's Sixth District Court of Appeal, conducted via Zoom and published on May 31, 2026, featured attorney Jerome Ramsaran acknowledging under direct questioning from three judges that the argument section of his appellate brief, prepared with the assistance of a retired lawyer on his staff, contained no legal argument at all. The document had the surface features of a professional legal filing, headers, citations, a statement of facts, but lacked the underlying analysis those features were meant to convey. The case is not cited in Lovett's paper, but it demonstrates the same gap the paper describes in the abstract: a document that passes surface validation while failing substantive validation, produced by someone who did not recognize the failure until it was named for him.

A comparable finding surfaced in software development, a field closer to the marketing technology stack PPC Land covers directly. Research from Anthropic involving 52 software engineers learning an unfamiliar Python library found that developers using AI assistance completed tasks marginally faster but scored 17 percent lower on subsequent comprehension tests, averaging 50 percent on knowledge quizzes covering concepts they had just applied, against 67 percent for those who coded manually. PPC Land reported on that research in February 2026, in coverage that quoted the study's authors warning that junior developers, under time and organizational pressure, may use AI "to complete tasks as fast as possible at the cost of skill development."

The economics behind the pattern

Lovett grounds the mechanism in labor economics dating to Gary Becker's 1964 distinction between firm-specific and general human capital. Firms have long underinvested in general training, the paper notes, because workers who gain broadly transferable skills can leave for a competitor before the training firm recovers its investment, a dynamic distinct from the upskilling debate over whether such investment deepens a role or merely reduces resistance to it. What changed with AI adoption, according to the paper, is not the underinvestment itself, which economists have documented for decades, but the disappearance of the productive activity through which junior training used to happen almost by accident. Organizations used to need entry-level workers to perform entry-level tasks; the tasks themselves were the training. Lovett writes that this arrangement was "not the product of governance; it was the product of accidental alignment," meaning firms maintained developmental pipelines because they needed the labor, not because they recognized any obligation to the profession. Artificial intelligence removes that operational necessity, and with it, the hidden mechanism that had been quietly regenerating expertise for decades.

The paper identifies five factors that determine how vulnerable a given profession is to this dynamic: how easily AI can substitute for the tasks that once trained novices, how strict regulatory requirements for demonstrated human expertise are in that field, how immediately visible the consequences of an expertise failure would be, how strong the profession's associations are at enforcing developmental standards, and how easily the work can be broken into discrete, AI-compatible pieces. Advertising and marketing sit closer to the high-vulnerability end of that scale than medicine or engineering, the paper implies, given the sector's relatively low regulatory intensity and highly modular workflows, though Lovett does not analyze the advertising sector by name.

What the paper recommends

The paper stops short of recommending that organizations reduce AI adoption. Its proposed remedies operate at three levels. At the organizational level, it points to practices such as phased AI introduction, in which developmental tasks are completed unaided before AI assistance is introduced, and structured workflows requiring active human reasoning before AI output is shown, an approach a randomized controlled trial by Sarah Everett and colleagues found preserved and even improved diagnostic accuracy among physicians using AI-assisted tools. At the professional-association level, Lovett draws on political economist Elinor Ostrom's research on how communities have historically sustained shared resources through mechanisms including monitoring, graduated sanctions, and collective decision-making, arguing that advertising and marketing associations with less-developed governance infrastructure may see commons depletion dynamics play out more acutely than fields such as medicine or law. At the policy level, the paper suggests options including training subsidies for organizations that preserve developmental pipelines, tiered credentials recognizing demonstrated independent expertise, and continuing education requirements that periodically test AI-free performance, framing these as hypotheses for further research rather than settled recommendations.

Lovett is careful throughout to distinguish what the evidence currently supports from what the framework predicts. The dissociation between AI-assisted performance and independent, unassisted capability is, in the paper's own words, "empirically established" wherever researchers have directly tested it. The claim that this dissociation will compound across an entire profession's workforce over ten to twenty years is described as "a structural prediction from that dissociation and commons theory, not an observed outcome," offered as a testable account of where current incentives lead rather than a claim that the depletion has already occurred broadly.

Why this matters for marketing and advertising

The paper does not analyze advertising, but its framework gives a name and a mechanism to something PPC Land's coverage has been documenting piecemeal across 2026: a hiring market in which AI fluency has become a baseline expectation for entry-level candidates even as the number of entry-level positions shrinks, and a workforce in which senior confidence in AI-generated outputs is rising even as the disciplined habit of checking those outputs, exactly the skill Lovett calls substantive validation, appears to be eroding among those who retain it. For an industry whose buying decisions increasingly run through agentic systems making autonomous choices about audience targeting, bid levels, and creative selection, the practical question the paper raises is not whether AI tools work. It is whether the people supervising those tools, five or ten years from now, will still have the depth of judgment needed to notice when they do not.

Timeline

  • 1964 - Gary Becker publishes his theory distinguishing firm-specific from general human capital, later applied by the paper to explain training underinvestment
  • 1968 - Garrett Hardin publishes "The Tragedy of the Commons" in Science, providing the metaphor the paper adapts
  • 1990 - Elinor Ostrom publishes "Governing the Commons," identifying design principles that let communities sustain shared resources through collective governance
  • October 2022 to September 2025 - Employment for workers aged 22 to 25 in AI-exposed US occupations falls 16 percent, according to Stanford Digital Economy Lab analysis of payroll data cited in the paper
  • November 2, 2025 - Nolan Lovett submits the manuscript to Human Resource Development Review
  • December 26, 2025 - Box chief executive Aaron Levie publishes an analysis arguing the Jevons paradox could mean AI expands total demand for marketing labor rather than reducing it, a case PPC Land covered
  • January 29, 2026 - Anthropic publishes research finding developers using AI assistance scored 17 percent lower on comprehension tests despite finishing tasks marginally faster, reported by PPC Land
  • April 20, 2026 - Basis publishes its 2026 Advertising Agency Report, finding 87.3 percent of agency professionals consider the traditional model broken and 39.9 percent of agencies reported layoffs in the prior year, covered by PPC Land
  • May 31, 2026 - A Florida Sixth District Court of Appeal hearing exposes an attorney's AI-linked brief containing no legal argument, illustrating the surface-versus-substantive validation gap the paper later describes; PPC Land reported on the hearing
  • June 2, 2026 - IAB Australia publishes its 2026 Talent Review, finding entry-level advertising vacancies had fallen to 1 percent of the available job pool, covered by PPC Land
  • July 6, 2026 - The manuscript is accepted for publication following two rounds of revision
  • July 26, 2026 - The paper is published online in Human Resource Development Review as an advance publication
  • July 31, 2026 - A version of the manuscript is posted to the arXiv preprint server

Summary

Who: Nolan Lovett, a researcher at NATO Special Operations University and doctoral candidate at Old Dominion University, authored the paper. It was published in Human Resource Development Review, a peer-reviewed academic journal.

What: The paper introduces the Cognitive Commons framework, arguing that AI adoption decisions which look efficient to individual organizations can collectively deplete the shared pool of deep professional expertise a field needs to renew itself, particularly the expertise required to catch errors in AI-generated work.

When: The manuscript was submitted November 2, 2025, accepted July 6, 2026, and published online in Human Resource Development Review on July 26, 2026. A version was posted to arXiv on July 31, 2026.

Where: The paper is a conceptual contribution to Human Resource Development scholarship, drawing on labor market evidence from the United States, United Kingdom, and Germany, with implications the author frames as applicable across professional sectors internationally rather than any single country.

Why: The paper matters to marketing and advertising because the sector shows several of the vulnerability factors the framework identifies, including low regulatory intensity, highly modular workflows, and, according to industry data PPC Land has separately reported, a sharp contraction in entry-level hiring even as senior reliance on unchecked AI output appears to be rising.