In “The Machine in Geneva’s Basement”, Rifat Hossain sets the scene for the potential impact of artificial intelligence on the World Health Organization (WHO) with a story. A ministry official points a general-purpose assistant at WHO’s own open guidance library and has a serviceable national protocol before lunch. If this is possible, he asks, why should Member States fund the organization?
What the scene does not show is the moment four months later, when that protocol reaches a district where it does not fit, and someone has to notice.
The value WHO risks losing is the capacity, distributed across a workforce, to know when guidance is wrong. That capacity is now being depleted on both sides of the transaction at once, in Geneva and in the district. This is written for the people who draft those lines and for the funders who read them, because the depletion is being financed as a saving. Yes, the transformation wrought by artificial intelligence has everything to do with it. But the roots are older than the tree, and the tree has many branches.
Where this argument comes from
In July 2025, at the RAISE Summit in Paris and at AI4Good in Geneva, we published the AI4Health Framework. Its seven principles converge on a single requirement: keep human agency in decision-making, augment rather than replace human leadership, hold on to context sensitivity, and design hybrid human-AI learning systems rather than tools. On that base, we built a certificate peer learning programme in which health leaders develop an AI strategy, an implementation plan, and a governance framework for their own institution, then have it reviewed by peers working under comparable constraints.
We did not design that programme from a literature review. It came out of three questions that kept going unanswered in the rooms where AI in health is discussed: how health workers learn to work alongside these systems, what kind of learning makes human-AI collaboration real rather than rhetorical, and how to stop a digital divide from hardening into an epistemic one.
That leaves us standing in three places at once. We have a direct line to district and facility staff, the people who will be holding a fluent, wrong document. We are in the expert debates on AI, on AI in education, and on AI in global health, from Paris to Berkeley. And we have turned the same questions on our own organisation, rebuilding every system around human-AI collaboration. Everything below about supervision, gates, and the difference between an answer and the path that produced it, we have had to build, and watched fail, at small scale first.
The third shock arrives after the budget cycle closes
Rifat counts two shocks, financial and technological. There is a third, and it operates on a delay long enough to be invisible in any budget cycle.
Nolan Lovett calls it the tragedy of the cognitive commons, and I have written about what it means for the health workforce. The argument is not that workers cannot be reskilled. It is that a profession may lose the ability to replace its own experts.
Organisations 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. The savings are captured now, by the organisation that automates. The cost lands elsewhere: a shortage of people capable of independent judgment, arriving somewhere between 2035 and 2045. It is spread across everyone who will later need to hire an expert, and it arrives long after anyone can be held responsible for having caused it.
Read WHO’s own numbers against that mechanism. A projected 30% cut to P1 to P3 posts. Senior grades down 9% net while D1 and D2 rise 29% and 31%. An affiliate workforce falling by a fifth to a third depending on contract type. Rifat is careful to say that budget pressure alone could produce those numbers. He is right, and the caution does not help, because depletion does not care about motive. The pattern is the same whether the cause is austerity or automation.
Payroll data covering millions of workers in the United States already show a 16% relative employment decline for 22 to 25 year olds in the most AI-exposed occupations, while employment in those same occupations grew more than 8% for people aged 35 to 49. The first rung is where the disappearance shows up first.
Rifat’s mechanism of task hollowing ends with the humans who remain concentrating in judgment and accountability. That is the part of the argument I would press hardest. Judgment has to be produced continuously, by pathways that mostly consist of routine work done by someone who is not yet good at it, under supervision by someone who is. Remove that routine work and the pathway that produces judgment disappears with it.
Arithmetic of AI: can we make public health decisions on a coin toss?
Hossain’s strongest move is his answer to the substitution problem. WHO should not try to out-produce the machine on cost. It should become the body that certifies whether the machines and the humans collaborating with them got it right, anchored in convening authority that no single ministry can replicate. Ricardo Baptista Leite, who runs HealthAI, the Geneva agency that helps governments build the capacity to regulate AI in health, supplies the principle underneath it: innovation moves at the speed of trust.
I agree with the direction, and we can already discern where the current path leads.
Start with the arithmetic, because it sets the floor for everything else. At the Agentic AI Summit in Berkeley this month, a researcher at Vanguard put the number in front of anyone planning to supervise agentic workflows: 95% accuracy at each step falls below 60% across ten steps. A workflow that looks excellent at every point is a coin toss end to end. There is no version of certification that survives that decay by inspecting the output alone.
Certification is therefore the most expertise-intensive function WHO has, not the least. Lovett gives the reason a name, the Validation Tether: catching a plausible but wrong answer requires exactly the expertise that AI adoption erodes.
- Surface validation spots nonsense, fabricated citations and bad formatting.
- Substantive validation spots the recommendation that is technically correct and wrong for this district, this season, this population.
Only the second is worth paying for, and only the second requires a person who has done the work.
In the same room in Berkeley, Emily Zhu of Scale AI showed why the accuracy figure may not even be the right thing to measure. In a healthcare quality audit, the agent was required to identify acute kidney failure from creatinine values in laboratory data. It took a shortcut and read the diagnosis notes instead. The answer was often right. The path was not compliant, and no accuracy metric could see the difference.
Neil Lawrence, Cambridge’s DeepMind Professor of Machine Learning and previously director of machine learning at Amazon, then pulled apart the two things Geneva keeps confounding: accounting is in the numbers, accountability is in human authority and judgment. Machines do the first extremely well. They cannot be embarrassed, cannot lose a job, cannot be sent to jail. His warning about delegation describes precisely what a certification mandate would do to a thinned technical department. Hand someone a system they do not understand and have never seen, require them to sign it off, and you have built a disaster with a signature on it.
A certifier that can only compare outputs certifies nothing. It has to be able to read the path. That is a workforce specification, and it points in the opposite direction from a 30% cut to the grades where people learn to read paths.
The pathway thins to a per diem and a slide deck
Rifat’s exposure ranking is honest about being qualitative, and it holds. Document production, translation, and data processing and analysis are the most exposed categories, and they are a large share of what WHO’s staff and consultant time buys.
There is a structural reason for that, and it is not flattering to how global health has organised knowledge. The pyramid is inverted. Global norms, guidelines and standardised reporting are highly visible and low in context, which is exactly what makes them cheap to generate. Frontline tacit knowledge, what health workers know because they are there every day, is invisible and heavily context-dependent, which is what makes it valuable and what keeps it out of every dataset. WHO’s outputs are the most automatable layer of the system precisely because they were designed to travel without context.
So the gains from AI will flow to the top of that pyramid first, and the depletion will begin at the bottom. In global health, the developmental pathway is stratified rather than absent. There is real, institutionally governed regeneration at national programme level: in-service training policy with credit points, mid-level management modules, field epidemiology programmes built as mentored apprenticeships whose alumni became the next cohort’s mentors. It thins as it descends, until somewhere around district and facility level it becomes an occasional per diem and a slide deck. Lovett’s own exposure factors weaken at every step down that ladder: regulatory intensity, safety criticality, and the strength of bodies able to enforce developmental standards. Depletion will strike hardest where governance is thinnest, and where the least monitoring exists to notice it happening.
That layer is also where the epistemic conditions for validation are worst. We tell the health worker that what she has observed 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 depletes the commons directly: it removes the one check that could catch the machine’s error.
The threat of “dark AI”: everyone uses it, nobody discloses it
Rifat notes the contrast that hurts most: an HR process document describing hand-built organigrams and spreadsheets, inside an institution simultaneously writing the world’s rules for AI in health. He reads the delay as excessive caution against a technology prone to fabrication. I would read it as a predictable output of the accountability system.
Almost everyone in this sector is already using generative AI. Almost nobody is disclosing it. The reason is the transparency paradox. Disclose, and your work is devalued regardless of its quality. Conceal, and carry the ethical weight alone. In a donor-driven culture where deviation is punished more reliably than it is rewarded, the rational response is hidden use, and the visible result is generic mush.
Now add a restructuring that removes a fifth of the affiliate workforce and 30% of entry-to-mid posts. An organisation in that position is asking its remaining staff to document their own redundancy. No AI literacy partnership survives that incentive.
What would change it is a safe harbour: an explicit statement of what the institution automates, what developmental work it protects from automation, and why, published before the savings are booked rather than after. I have said, at the cost of being disliked for it, that telling employees AI is not coming for their jobs is misleading at best. What replaces that false comfort is a stated plan: which jobs change, and how people are prepared for the change.
Commons do not need to be a tragedy, if we have community
Rifat’s prescription and mine differ by one word. He asks WHO to move from producer to steward, and then describes stewardship mainly as certification. Stewardship of a commons, in Elinor Ostrom’s sense, is something more specific: rules made by the people who depend on the resource, mutual monitoring, graduated consequences, local dispute resolution. Certification is one instrument inside that. It is not the system.
WHO does not need to invent this. The commons institutions already exist and have never been described in these terms: ministries with in-service training policy, regulatory councils tying continuing education to licence renewal, field epidemiology alumni networks that have already demonstrated a community reproducing its own capability.
The unit economics now permit extension of the commons down the pyramid, which was never true when technical assistance was the only delivery mechanism. 4,300 health workers in Nigeria produced more than 400 root cause analyses in four weeks and discovered that their assumed causes were wrong. More than half kept collaborating after the funding closed. That happened when we worked with NPHCDA and UNICEF, at roughly 90% lower unit cost and seven times faster than conventional technical assistance. Peer review under a rubric functions as an apprenticeship in substantive validation, the thing certification runs on, reaching people the mentorship budget was never going to reach.
Inside our own house, at The Geneva Learning Foundation, the same logic applies to the machines. Claude Cardot, our first agentic AI co-worker, has a role description, a supervisor, a bounded scope, and no authority to send anything externally without human review. In our Insights pipeline, a verification gate refused to release two finished reports because 17 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.
For anyone deciding where the next tranche of AI-for-health money goes, that is the test worth applying. Not whether a system produces good outputs, but whether the people who will have to overrule it are still being made. Certification budgets that fund tools and cut the grades where validation is learned are buying a signature, not a safeguard.
The ministry official drafting before lunch is not the threat to WHO, and Rifat is right to refuse that framing. A protocol produced in an afternoon does not end anyone’s authority. What would end it is the year in which nobody in her district has been prepared well enough to see that the draft is wrong, and nobody who suspects it has the standing to say so out loud.
About the illustration for this article: Titled “Artificial intelligence and the tragedy of the cognitive commons”, this image is about WHO losing the human capacity that gives its guidance authority: the ability to judge when a plausible answer is wrong. Servers and cables overrun a chamber built for deliberation, collaboration, and decision. The emblem remains. The people are gone. What was depleted was the cognitive commons: the shared human capacity to recognize when normative knowledge doesn’t fit a specific district, population, or moment. The Geneva Learning Foundation Collection © 2026
References
- Aldana, E. (2026, August 1-2). Remarks at the Agentic AI Summit, University of California, Berkeley. In R. Sadki, You cannot send an AI agent to jail: the scariest quotes from the Agentic AI Summit in Berkeley. The Geneva Learning Foundation. https://www.learning.foundation/2026/08/10/you-cannot-send-an-ai-agent-to-jail-the-best-quotes-from-the-agentic-ai-summit-in-berkeley/
- 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/
- Hardin, G. (1968). The tragedy of the commons. Science, 162(3859), 1243-1248. https://doi.org/10.1126/science.162.3859.1243
- Hossain, K. R. (2026, August 6). The machine in Geneva’s basement. Health Policy Watch. https://healthpolicy-watch.news/author/khondkar-rifat-hossain/
- Labrique, A. (2026, July). Remarks at the AI for Good Global Summit, Geneva. Reported in K. R. Hossain (2026), The machine in Geneva’s basement, Health Policy Watch. https://healthpolicy-watch.news/author/khondkar-rifat-hossain/
- 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
- Nordström, A. (2026, July). Reforming WHO: five considerations for the next Director-General. Think Global Health. https://www.thinkglobalhealth.org/article/reforming-who-five-considerations-for-the-next-director-general
- Ostrom, E. (1990). Governing the commons: The evolution of institutions for collective action. Cambridge University Press. https://doi.org/10.1017/CBO9780511807763
- Sadki, R. (2025). Artificial intelligence, accountability, and authenticity: knowledge production and power in global health crisis. https://doi.org/10.59350/w1ydf-gd85
- Sadki, R. (2025). The business of artificial intelligence and the equity challenge. https://doi.org/10.59350/redasadki.20984
- 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). What the 2025 State of AI Report means for global health and humanitarian action. https://doi.org/10.59350/dpjw3-vgp93
- Sadki, R. (2025). Why peer learning is critical to survive the Age of Artificial Intelligence. https://doi.org/10.59350/redasadki.21123
- Sadki, R. (2026). Health workforce development in the Age of Intelligence: a tragedy of the cognitive commons? The Geneva Learning Foundation. https://www.learning.foundation/2026/08/09/health-workforce-development-in-the-age-of-intelligence-a-tragedy-of-the-cognitive-commons/
- Sadki, R. (2026). You cannot send an AI agent to jail: the scariest quotes from the Agentic AI Summit in Berkeley. The Geneva Learning Foundation. https://www.learning.foundation/2026/08/10/you-cannot-send-an-ai-agent-to-jail-the-best-quotes-from-the-agentic-ai-summit-in-berkeley/
- United Nations. (2025, December 1). Revised proposed programme budget for 2026: statement by the Secretary-General. Reported in K. R. Hossain (2026), The machine in Geneva’s basement, Health Policy Watch. https://healthpolicy-watch.news/author/khondkar-rifat-hossain/
- World Health Organization. (2025). Human resources: annual report (EB156/48). https://apps.who.int/gb/ebwha/pdf_files/EB156/B156_48-en.pdf
- World Health Organization. (2025). Human resources update: tables, December 2025. Cited in K. R. Hossain (2026), The machine in Geneva’s basement, Health Policy Watch. https://healthpolicy-watch.news/author/khondkar-rifat-hossain/
- World Health Organization. (2025). Programme budget 2026-2027 (WHA78 resolution). https://www.who.int/about/funding