Two people, both of whom I have met versions of many times.
The first is a national programme manager at Nigeria’s National Primary Health Care Development Agency (NPHCDA), or at the Ghana Health Service.
She is formidable.
She has been through in-service training with credit points that counted toward her career progression, mid-level management modules, a field epidemiology programme where a designated mentor reviewed her outbreak investigation and sent it back for revision, and 15 years inside an institution with a memory.
She did not make it on her own, and she would be the first to say so.
Her organization has real frailties, chronic underfunding, cascade training that thins as it descends, mentorship budgets that vanish first when a grant closes.
It is still an organizational culture that formed her, and pretending otherwise insults both her and the colleagues who invested in her.
The second is a district immunization officer three administrative layers below.
She attended a two-day cascade workshop in the state capital four years ago, received a per diem and a slide deck, and then spent a decade figuring out alone why the children in three settlements kept being missed.
She has real expertise too.
Nobody designed the pathway that produced it.
Hold both of them in mind, because the difference between them is the whole argument of this article.
What is the ‘tragedy of the cognitive commons’?
Nolan Lovett’s “The tragedy of the cognitive commons: How AI could disrupt the regeneration of professional expertise”, published in Human Resource Development Review, lays out a fascinating argument about artificial intelligence and expertise.
Its claim is that we have been asking the wrong question.
The question is not whether workers can be reskilled.
The question is whether a profession can still replace its own experts.
Two old ideas, in plain language
The paper builds on two pieces of social science that are worth explaining, because the argument collapses without them.
The first is a parable.
In 1968, the biologist Garrett Hardin described a village pasture open to all.
Every herder gains personally from adding one more animal, and the cost of the extra grazing is spread across everybody.
Each decision is individually sensible.
Collectively they strip the pasture bare.
He called it the “tragedy of the commons”, and it became the standard way of describing how a shared resource can be destroyed by people who are each behaving reasonably.
Lovett is explicit that he borrows the shape of Hardin’s story and rejects its pessimism.
Which brings in the second idea, and the more hopeful one.
Elinor Ostrom won the Nobel Prize in economics for demonstrating that Hardin was empirically wrong about the ending.
She spent decades documenting real communities, irrigation systems in Spain and the Philippines, mountain pastures in Switzerland, Japanese village forests, that had shared a finite resource for centuries without destroying it.
They did it by making rules together: agreeing who has access, monitoring each other’s use, imposing modest and graduated consequences on people who take too much, and resolving disputes locally.
Crucially, Ostrom found these arrangements were almost never imposed from outside.
They were invented and adapted by the people who depended on the resource.
So when I say the paper follows Ostrom to the governance question, this is what I mean.
The tragedy is not a prophecy.
It is a diagnosis of what happens in the absence of institutions, which turns the analysis into a practical question: who makes the rules that keep this particular pasture alive, and at what level?
What is being grazed
The resource in Lovett’s account is what he calls the Cognitive Commons: the pool of practitioners in a field who hold deep domain knowledge, tacit understanding, and the judgment to work independently when a situation falls outside the protocol.
It is not human capital, which sits inside a person or a firm.
It is not a community of practice, which is a social process rather than a stock.
It is not workforce capacity, which recruitment can move around but cannot manufacture.
From there the paper makes four moves.
It separates Internalized Mastery, built through sustained cognitive struggle, from Distributed Mastery, the fluency of orchestrating AI systems.
It names the Validation Tether: catching a plausible but wrong answer requires the very expertise that AI adoption may be eroding.
Surface validation spots nonsense and bad formatting. Substantive validation spots the recommendation that is technically correct and wrong for this patient, this district, this season.
It identifies the mechanism as an accident that has now stopped happening. Organizations never maintained developmental pipelines out of stewardship. They hired juniors because juniors did the junior work, and expertise regenerated as a by-product.
AI severs the two.
And it prices the harm.
An organization that eliminates entry-level roles keeps all the savings and spreads the depletion cost across everyone who will later need to hire an expert.
The delay hides it: decisions taken in 2023 surface as expert scarcity around 2035 to 2045, long after anyone can be held responsible.
The evidence is early, and the paper says so.
Payroll data covering millions of U.S. workers show a 16% relative employment decline for 22 to 25 year olds in the most AI-exposed occupations from October 2022 to September 2025, while employment for 35 to 49 year olds in those same occupations grew more than 8%.
Endoscopists lost independent detection accuracy after adopting AI assistance.
Consultants improved on tasks inside the AI’s competence and got worse on tasks outside it, exactly where judgment matters.
In one experiment, more than 80% of participants noticed errors in AI recommendations and followed them anyway.
Lovett grades his own claims into established findings, theoretical derivations, and predictions requiring test.
That discipline is worth imitating, and I will try.
Where this describes our world exactly
Three of his findings land directly on what we have been documenting.
The performance-learning gap is the gap the OECD reported at Digital Education Outlook 2026: students who wrote better essays with a language model and could not afterwards remember what they had written about, students who completed more mathematics exercises and understood less mathematics.
Dragan Gasevic called it metacognitive laziness.
Jason Lodge called it an e-bike used to get fit.
Lovett calls it the dissociation between assisted performance and independent capability.
Three vocabularies, one phenomenon.
The disappearing first rung is what Matt Sigelman brought to the Empower Learners conference and what Beatriz Moutinho of Cabo Verde named more precisely than any economist when she said she feared young people preemptively replacing themselves.
And the Validation Tether is the risk I have described as the danger not that a health worker uses AI to draft a report, but that she trusts its conclusions without the validation that prevents catastrophic error, and that when capital investment outruns human activation, the return sits on foundations that will not hold.
If the World Bank’s jobs doctrine rests on infrastructure, human capital, and catalytic finance, Lovett has just told us the second pillar is not a stock to be built but a flow to be governed.
The commons is stratified, not absent
Lovett states his boundary condition.
His framework assumes a model of professional formation in which expertise develops through entry-level employment inside hierarchical organizations and is governed by formal credentialing, and he notes that this model dominates Western professional sectors, without exhausting the ways expertise has been transmitted.
My first instinct was to say that in our corner of the universe – global health, global development, and humanitarian response – this pathway does not exist.
That instinct was wrong, because a romantic version of the frontline erases the institutions that actually built the leaders I most admire.
The Ghana Health Service has in-service training policies specifying training needs, frequency, curricula, and credit points that count toward career progression, alongside professional regulatory bodies running continuing education tied to licence renewal.
Nigeria’s national immunization strategy has for years organized mid-level management training and integrated modules for managers and service providers.
Field epidemiology training programmes across Africa are explicitly built as two-year mentored apprenticeships, learning by doing with a designated mentor and a target ratio of one mentor to five residents, and their alumni networks have become the mentor pool for later cohorts.
When the third Zambian cohort needed mentors, the first cohort was called on and answered.
That is a regeneration mechanism.
It is recognizably the thing Lovett is describing, and it exists in Accra, Abuja, Abidjan, or Ouagadougou.
It is also, in the same breath, fragile in documented ways.
The 2024 needs assessment of Ghana’s own advanced field epidemiology programme found limited mentorship funding, low programme visibility, and insufficient funding for resident development.
National strategies still describe retraining frontline workers in a cascaded manner, and everyone who has worked inside a cascade knows what reaches the fifth iteration.
So the accurate claim is not that the commons was never built.
It is that the commons is stratified.
There is a genuine, institutionally governed regeneration pathway at national and sub-national programme level, thinning as it descends, until somewhere around the district and facility level it becomes a per diem and a slide deck.
This matters for three reasons.
First, it corrects the politics.
The people who benefited from those pathways are not exceptions who transcended a broken system.
They are evidence that deliberate workforce development works, which is the strongest available argument for extending it rather than replacing it.
Second, it locates the vulnerability precisely.
Lovett offers five factors that determine how exposed an occupation is, among them regulatory intensity, safety criticality, and the strength of professional bodies able to enforce developmental standards.
Read down our pyramid and every one of those protections weakens as you descend.
The national programme manager sits inside licensure, mentorship, and an alumni network.
The district officer may sit inside some or none of them.
If the framework predicts that depletion strikes hardest where governance is thinnest, then in global health it will strike from the bottom up, and it will strike the layer where the least monitoring exists to notice.
Third, it identifies the actors.
Lovett looks for commons stewards and finds professional associations in medicine, law, and engineering, while observing that the most exposed sectors have the weakest capacity for this.
In our sector the candidates already exist and are rarely described in these terms: ministries of health with in-service training policy, regulatory councils with continuing education requirements, field epidemiology alumni networks that have already demonstrated the Ostrom pattern of a community reproducing its own capability.
They are commons institutions, and nobody has told them so.
The strategic question therefore is not how to invent stewardship from nothing.
It is whether those institutions will extend stewardship down the pyramid, in time, and at a cost per practitioner that a national budget can carry.
The variable the framework is missing
There is one more thing, and it is the piece I care about most.
Lovett’s own analysis makes substantive validation depend on two conditions: the domain knowledge that makes an error recognizable, and the epistemic stance that prompts a practitioner to look in the first place.
He is right, and he cites the research on epistemic beliefs to say so.
Now consider what our sector does to that stance.
We tell the frontline health worker that what she knows is merely anecdotal.
We place practitioner experience at the bottom of the hierarchy of evidence.
We ask how many children she vaccinated and almost never what she noticed about them.
A workforce systematically taught that its own observation does not count as knowledge will not challenge a fluent machine.
Epistemic injustice is not only a moral failure.
It is a mechanism of commons depletion, because it degrades the Validation Tether directly: it trains the only people positioned to catch a contextually wrong answer to defer to whatever arrives carrying institutional authority.
That variable is not in the framework.
It should be.
And it compounds the stratification, because the deference is steepest exactly where the developmental pathway is thinnest.
What follows for our work
If the accidental mechanism is breaking, and the deliberate one reaches only the upper layers, then regeneration has to become designed, extendable, and cheap enough to run at the scale of the actual workforce.
That is a description of what we have been building for a decade at The Geneva Learning Foundation (TGLF), and the paper gives a sharper account of why it works.
Cognitive struggle is the product, not the friction. Lovett is explicit that AI reallocates cognitive work rather than simply removing it, and that assistance which preserves the practitioner’s own attempt and requires her to evaluate, question, and justify can itself be developmental. A randomised trial of clinician-AI collaboration found the same: workflows requiring active engagement with AI reasoning preserved and even improved diagnostic performance. This is why our questions are inverted. We do not ask what should be done about vaccine hesitancy. We ask her to describe a time she addressed it, what she did, and what happened. A model can answer the first question. Only someone who was there can answer the second, and answering it is itself an act of learning.
Peer review is a regeneration mechanism, not a cheap substitute for expert grading. Applying a rubric to three colleagues’ work is how evaluative judgment forms, and evaluative judgment is precisely what the Validation Tether requires. TGLF’s sixteen-day peer learning exercise is a highly-condensed cycle of development, review, and revision. It is an apprenticeship in substantive validation, each time reaching thousands of people the mentorship budget was never going to reach.
Praxis closure consolidates it. Monday’s commitment to one concrete action, midweek comparison with peers, Friday’s report including failure, next Monday’s goal set with that experience behind it. That is the rhythm set by TGLF’s Impact Accelerator. Lovett’s developmental account is scaffolded engagement with progressively complex problems in which practitioners struggle, monitor their reasoning, and correct flawed assumptions against evidence. That is the Accelerator loop, written in the vocabulary of human resource development.
The persistent network is a governance layer, and it should be built with the existing ones rather than beside them. Ostrom’s finding was that durable arrangements emerge from local experimentation, with shared standards, mutual monitoring, and conflict resolution. Rubrics are standards. Peer review is monitoring. Reciprocity is what keeps contribution rational. When 4,300 health workers in Nigeria produced more than 400 root cause analyses in four weeks, discovered their assumed causes were wrong, and more than half kept collaborating after the funding closed, at roughly 90 percent lower unit cost and seven times faster than conventional technical assistance, that happened in partnership with NPHCDA and UNICEF. Not as an alternative to the national system, but as the layer the national system could not otherwise afford to reach.
AI belongs inside the structure, tethered. Our first agentic AI hire, Claude Cardot, has a role description, a supervisor, a bounded scope, and no authority to send anything externally without human review. That is the Validation Tether written into an org chart. The same logic runs through our Insights pipeline, where a verification gate refused to release two finished reports because seventeen quotations could not be matched character by character against the raw exports. Gates, not guidelines. A standard that lives as an instruction to a model will eventually be self-reported as satisfied.
What we should be honest about
The framework indicts us too.
When Joseph Ngugi, a diligent Kenyan Scholar, began submitting generic AI-generated prose, that was not disengagement. (Joseph passed away last year, and I miss him.)
The transparency trap explains it: disclose AI use and risk having your knowledge dismissed, conceal it and carry the ethical weight alone.
If our evidence rests on first-person practitioner accounts, and those accounts are increasingly smoothed by a model before they reach us, then the foundation of everything we claim through our AI work is exposed to precisely the augmentation-without-internalization mechanism Lovett describes.
Prohibition will not fix that.
Only conditions in which transparency is safe will.
I have also said, at the cost of being disliked for it, that telling employees AI is not coming for their jobs is misleading at best.
An organization arguing for commons stewardship while quietly free-riding is not credible.
What the paper obliges us to do is say publicly what we automate, what developmental work we protect from automation, and why.
There is a measurement obligation too.
Lovett asks for indicators of commons health rather than training satisfaction: does a cohort develop independent capability, and how fast.
We are unusually well placed to answer, because we measure value creation against a baseline drawn from more than 10,000 health workers, and because our reliability protocol publishes its failures.
15 of 19 codes fell below the kappa floor in our last validation run, and the report says so in its methods section.
If we want percentages to carry their own epistemology, our claims about expertise formation should carry theirs.
The question that remains
Lovett ends by asking whether his field will engage this or stay confined by assumptions that no longer describe the world.
I would put a different question to ours.
The San Francisco Consensus holds that within three to six years AI will reason at expert levels, coordinate complex work as agents, and understand any request in any language.
If that is even approximately right, the binding constraint on global health will not be access to intelligence.
It will be whether enough people retain the internalized judgment to know when the intelligence is wrong, and the standing, safety, and structure to say so out loud.
Hardin’s pasture was destroyed because nobody owned the consequences.
Ostrom’s villages survived because the people who depended on the resource made the rules together.
The institutions that formed the national programme manager already know how to do this, at their level, imperfectly, under-funded, and for a fraction of the workforce.
The work now is to extend that pathway to the district officer before the accident that used to produce her expertise stops happening altogether.
The tragedy is not inevitable.
But it will not be averted by better training, and it will not be averted at all while we treat the knowledge of the people who are there every day as the cheapest thing in the system.
References
- 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. https://doi.org/10.1177/15344843261470602
- Hardin, G. (1968). The tragedy of the commons. Science, 162(3859), 1243-1248.
- Ostrom, E. (1990). Governing the commons: The evolution of institutions for collective action. Cambridge University Press.
- Ostrom, E. (2010). Beyond markets and states: Polycentric governance of complex economic systems. American Economic Review, 100(3), 641-672. https://doi.org/10.1257/aer.100.3.641
- Becker, G. S. (1964). Human capital: A theoretical and empirical analysis, with special reference to education. National Bureau of Economic Research.
- Brynjolfsson, E., Chandar, B., & Chen, R. (2025). Canaries in the coal mine? Six facts about the recent employment effects of artificial intelligence. Stanford Digital Economy Lab. https://digitaleconomy.stanford.edu/publications/canaries-in-the-coal-mine/
- Hampole, M., Papanikolaou, D., Schmidt, L. D. W., & Seegmiller, B. (2025). Artificial intelligence and the labor market (Working Paper No. 33509). National Bureau of Economic Research. https://doi.org/10.3386/w33509
- Budzyń, K., Romańczyk, M., Kitala, D., et al. (2025). Endoscopist deskilling risk after exposure to artificial intelligence in colonoscopy: A multicentre, observational study. The Lancet Gastroenterology & Hepatology, 10(10), 896-903. https://doi.org/10.1016/S2468-1253(25)00133-5
- Vicente, L., & Matute, H. (2023). Humans inherit artificial intelligence biases. Scientific Reports, 13(1), 15737. https://doi.org/10.1038/s41598-023-42384-8
- Dell’Acqua, F., McFowland, E. III, Mollick, E., et al. (2026). Navigating the jagged technological frontier: Field experimental evidence of the effects of artificial intelligence on knowledge worker productivity and quality. Organization Science, 37(2), 403-423. https://doi.org/10.1287/orsc.2025.21838
- Wiles, E., Krayer, L., Abbadi, M., et al. (2024). GenAI as an exoskeleton: Experimental evidence on knowledge workers using GenAI on new skills. SSRN. https://doi.org/10.2139/ssrn.4944588
- Everett, S. S., Bunning, B. J., Jain, P., et al. (2025). From tool to teammate: A randomized controlled trial of clinician-AI collaborative workflows for diagnosis. medRxiv. https://doi.org/10.1101/2025.06.07.25329176
- Bainbridge, L. (1983). Ironies of automation. Automatica, 19(6), 775-779. https://doi.org/10.1016/0005-1098(83)90046-8
- Lave, J., & Wenger, E. (1991). Situated learning: Legitimate peripheral participation. Cambridge University Press.
- Ericsson, K. A. (2006). The influence of experience and deliberate practice on the development of superior expert performance. In The Cambridge handbook of expertise and expert performance (pp. 683-703). Cambridge University Press.
- Eraut, M. (2004). Informal learning in the workplace. Studies in Continuing Education, 26(2), 247-273. https://doi.org/10.1080/158037042000225245
- Hutchins, E. (1995). Cognition in the wild. MIT Press.
- Polanyi, M. (1966). The tacit dimension. Doubleday.
- Fricker, M. (2007). Epistemic injustice: Power and the ethics of knowing. Oxford University Press.
- Ministry of Health, Republic of Ghana. Human resource for health policies and strategies. https://www.moh.gov.gh/wp-content/uploads/2016/02/Human-Resource-Health-Policy-Plan.pdf
- Federal Republic of Nigeria. Comprehensive EPI multi-year plan. https://www.nitag-resource.org/sites/default/files/2b302ead55915d61468ac747e2e437aee65871b1_1.pdf
- André, A. M., Lopez, A., Perkins, S., et al. (2017). Frontline field epidemiology training programs as a strategy to improve disease surveillance and response. Emerging Infectious Diseases, 23(Suppl 1). https://pmc.ncbi.nlm.nih.gov/articles/PMC5711307/
- Needs assessment of the advanced Ghana Field Epidemiology and Laboratory Training Program, April 2024: lessons learned and best practices. (2025). https://pubmed.ncbi.nlm.nih.gov/41409400/
- Zambia Field Epidemiology Network and Zambia National Public Health Institute (2023). Strengthening the public health workforce for Africa through field epidemiology training programmes. 3rd International Conference on Public Health in Africa. https://www.youtube.com/watch?v=B7ICGbwomrI
- Wenger, E., Trayner, B., & de Laat, M. (2011). Promoting and assessing value creation in communities and networks: A conceptual framework. Ruud de Moor Centrum, Open Universiteit.
- Sadki, R. (2026). Research in the Age of Artificial Intelligence: early learning from our insights pipeline. https://www.learning.foundation/2026/07/29/research-in-the-age-of-artificial-intelligence-early-learning-from-our-insights-pipeline/
- Sadki, R. (2026). Future of the workforce: the 780 million job gap, the AI reckoning, and the missing link in the World Bank’s biggest bet. https://doi.org/10.59350/8mk5d-59c17
- Sadki, R. (2026). AI self-replacement: what happens when we delegate our thoughts to artificial intelligence? https://www.learning.foundation/2026/04/03/ai-self-replacement-what-happens-when-we-delegate-our-thoughts-to-artificial-intelligence/
- Sadki, R. (2026). OECD Digital Education Outlook 2026: how can AI help human beings learn and grow? https://www.learning.foundation/2026/03/24/oecd-digital-education-outlook-2026-how-can-ai-help-human-beings-learn-and-grow/
- Sadki, R. (2026). Introducing Claude Cardot, our first AI co-worker to support frontline health and humanitarian leaders. https://www.learning.foundation/2026/03/13/introducing-claude-cardot-our-first-ai-co-worker-to-support-frontline-health-and-humanitarian-leaders/
- Sadki, R. (2025). The great unlearning: notes on the Empower Learners for the Age of AI conference. https://doi.org/10.59350/859ed-e8148
- Sadki, R. (2025). The crisis in scientific publishing: from AI fraud to epistemic justice. https://doi.org/10.59350/532vk-hq503
- Sadki, R. (2025). Why peer learning is critical to survive the Age of Artificial Intelligence. https://doi.org/10.59350/redasadki.21123
- Sadki, R. (2025). The agentic AI revolution: what does it mean for workforce development? https://doi.org/10.59350/redasadki.21092
- Sadki, R. (2025). What does AI reasoning mean for global health? https://doi.org/10.59350/redasadki.21102
- Sadki, R. (2025). Language as AI’s universal interface: what it means and why it matters. https://doi.org/10.59350/redasadki.21088
- Sadki, R. (2025). Eric Schmidt’s San Francisco Consensus about the impact of artificial intelligence. https://doi.org/10.59350/redasadki.21109
- Sadki, R. (2025). A global health framework for artificial intelligence as co-worker to support networked learning and local action. https://doi.org/10.59350/gr56c-cdd51
- Sadki, R. (2025). The business of artificial intelligence and the equity challenge. https://doi.org/10.59350/redasadki.20984