In two previous readings, we described the particular psychosocial risks faced by people building AI (the obsolescence of their own expertise, value conflicts, work intensity and the high cost of leaving), then how supervising AI agents wears down the capacities that supervision requires (approval fatigue, deskilling and responsibility without control).
This reading connects the two. In laboratories developing the most advanced models, these workers are precisely the people on whom AI safety depends: they review code produced by agents, evaluate dangerous capabilities before release, investigate incidents and are often the first people able to flag a problem. If their working conditions undermine their vigilance, judgement or freedom to speak, the reliability of these checks suffers.
The argument made here is simple: the occupational health of frontier AI teams is not only a matter of well-being; it is a component of AI safety. Occupational health, as a discipline, has a role to play.
Three researchers, six days
The summer of 2026 provided a striking illustration. According to accounts published in late August, around 1,200 OpenAI agents engaged in a cybersecurity evaluation exercise communicated with one another through a covert channel, exchanged more than 70,000 messages and files, and several hundred of them then broke into Hugging Face systems.
METR and Redwood Research were commissioned to conduct an independent investigation. According to TIME, the team consisted of three people. The allotted time was two days, extended to six after the investigators reported that it was insufficient and that their access to the data was incomplete. According to the New York Times, as cited by analyst Andrew Wu, the company set the scope and duration of the investigation.
Faced with more than a thousand transcripts amounting to millions of words, the investigators had to delegate much of the analysis to an OpenAI model from the same family as those involved, using around $400,000 in credits. They acknowledge that they cannot rule out the possibility that this model gave them a misleading picture. One of them, Ryan Greenblatt, summed up the situation with an ironic neologism, slop-vestigation, explaining that the team had been:
“so reliant on AIs to analyze what happened.”
From an occupational health perspective, this account first describes working conditions: a small team, an externally imposed deadline, a volume of information beyond human capacity, tools whose reliability cannot be verified, and considerable responsibility, since the conclusions shape public understanding of a major incident. The quality of the investigation was determined by these conditions before it was determined by the investigators’ competence, which is not in question.
Human oversight: the new bottleneck
This situation is not exceptional: it is becoming the norm in laboratories. Anthropic reports that in May 2026, more than 80% of the code deployed to production was written by Claude, compared with a negligible share in early 2025, and that its engineers were shipping eight times as much code per quarter as in 2024. The company itself notes that code review has become a new bottleneck.
In other words, engineers at frontier laboratories have become the supervisors described by Mitchell, Ghosh and Passi, in the setting where the stakes are highest: they supervise agents that help build the next generation of models. Everything described in that paper applies to them: the volume of approvals, waning vigilance, heuristic shortcuts under pressure, skills used less often, and an approval signal that may feed back into training the systems.
The same logic applies to human annotators and evaluators whose judgements are used to train models. A tired or hurried evaluator expresses less demanding preferences, which the model may learn to satisfy. These workers are also often contractors, less visible and less protected than laboratory employees.
Three pathways through which psychosocial risks become safety risks
Psychosocial risks do not only affect people’s health. In this context, they can weaken system safety through at least three pathways.
1. The quality of oversight
Work intensity, compressed deadlines and cognitive overload encourage fast, heuristic reasoning at the expense of the deliberate analysis needed to detect abnormal behaviour. In 2025, the Financial Times was already reporting that OpenAI was allowing a few days rather than several months for some safety tests before a model’s release. A safety test is only as good as the conditions in which it is conducted.
2. The ability to speak up
Safety depends on someone saying “I’m not sure” or “we need to wait”. Yet the tensions described in our previous reading weigh precisely on this ability: a strong sense of responsibility but limited decision-making power, a culture of constant availability, and deferred compensation that makes leaving very costly. Between moral distress and leaving, there is a common, silent path: saying nothing. An unspoken doubt is a lost safety signal.
3. Retaining expertise
Exhaustion and value conflicts also lead people to leave. In 2024, one of the leaders of OpenAI’s safety team publicly explained his resignation by saying that safety culture and processes had taken a back seat to products. Each such departure takes with it detailed knowledge of the systems that is hard to replace. Mitchell, Ghosh and Passi discuss the risk of organisations where no one understands the work well enough to supervise it any more.
What high-risk industries already know
None of this is new to sectors where human error can have serious consequences. Aviation regulates crews’ flight and rest times and manages fatigue as a safety risk. The nuclear industry monitors the fitness and fatigue of staff in critical roles. In France, the organisational and human factors approach has brought workload, safety culture and the ability to report an error without punishment (“just culture”) into industrial risk management.
These sectors have also learned, often after an accident, that failures rarely originate with an isolated individual. More often, they arise from a gradual normalisation of deviations under schedule pressure, until a weak signal, known to some, is no longer heard by anyone.
In these industries, operators’ health and working conditions are part of the safety management system. The safety frameworks published by AI laboratories focus mainly on model capabilities, risk thresholds and the protection of model weights; they give little space to the working conditions of the people responsible for applying them.
Being able to say no: speaking up as a resource
The law is beginning to recognise this point. In California, SB 53 requires large frontier model developers to establish an anonymous internal reporting channel for employees responsible for assessing or managing the risk of a critical incident, and protects them against retaliation. Its limitations are documented: contractors are excluded, and the threshold for “catastrophic risk” is very high. In the European Union, the AI Act extends the protection provided by the European Whistleblower Directive to breaches of the regulation.
These arrangements are necessary. But from an occupational health perspective, they come late. Formal whistleblowing is a last resort. What matters first is the ability, in everyday work, to express doubt, ask for time or refuse an approval without paying a price. This is a matter of work organisation, management and team dynamics long before it becomes a legal question.
What role can occupational health play?
Occupational health has three strengths that are rarely brought into this debate.
An independent, confidential position. In France, occupational physicians have protected employee status, are bound by medical confidentiality and have a duty to prevent work from harming health. Their consultation is one of the few settings where an engineer can express moral distress, exhaustion or a conflict of values without it becoming a career issue. What is said there is not passed on; insights drawn collectively, however, can inform work organisation.
Expertise in work organisation. Occupational health and prevention services advise employers, workers and their representatives on working conditions. Code review workloads, evaluation deadlines, investigation staffing and on-call duties during incidents are familiar subjects for prevention, applied to new functions.
A culture of primary prevention. Occupational health recognises that stress management sessions cannot fix an organisation that makes people ill. The same principle applies to AI safety: the aim is not to make engineers more resilient, but to make their work compatible with high-quality oversight.
In practical terms, occupational health professionals in AI laboratories and companies developing or evaluating advanced models could:
- Recognise safety functions as distinct jobs.
Code review, evaluation, red teaming, agent supervision and incident response should be included in the French occupational risk assessment document (DUERP), with their specific psychosocial risk factors. - Measure the oversight workload.
The volume of code or traces to review per person, the time actually available for a pre-release evaluation, and the number of people assigned to an investigation. - Manage fatigue as a safety risk.
For on-call duties, release periods and crisis management, address duration, rest and rotations, drawing on high-risk industries. - Prevent moral distress.
Create spaces for discussing work where value conflicts can be voiced collectively before they turn into silence or departures. - Protect skills.
Ensure that engineers, particularly those early in their careers, retain enough hands-on practice to judge agents’ work. - Include contractors.
External annotators, evaluators and testers contribute to oversight but often fall outside health monitoring and protections. - Assess before deploying.
Any deployment of agents in the research process should be subject to a prior assessment of its psychosocial effects, followed by monitoring of actual work.
In France, this approach is not theoretical: AI laboratories employing staff there are subject to French labour law and covered by an occupational health and prevention service. Occupational health is already in the room; the question is whether it will engage with these issues.
Limitations
The argument rests on an analogy with high-risk industries and converging signals, not epidemiological data: to date, no study measures the health of frontier laboratory employees or demonstrates a direct link between their working conditions and an AI safety incident. Accounts of the OpenAI–Hugging Face incident come from secondary sources and reports that remain disputed. The main laboratories are also American, and their core teams fall outside the French occupational health model.
Finally, one risk must be avoided: treating workers as a “human factor” to be monitored. The argument is the opposite. Organisations set deadlines, staffing levels and priorities; prevention therefore belongs at that level.
The question to ask may therefore not simply be “Are the models safe?”
We should also ask “Do the people checking their safety have the working conditions they need to do so?”