Protect workers.
Identify how AI changes workload, autonomy, skills and working relationships. Involving the people affected helps compare promised gains with actual constraints and develop preventive measures before difficulties become established.
AI safety aims to prevent the harm that artificial intelligence systems can cause to people and society.
At work, this concerns health, autonomy, skills and the ability to act. This initiative proposes that occupational health contribute to the design, evaluation and deployment of safer AI.
Safety depends on models and on their interactions with people and organisations. The sociotechnical approach proposed by Weidinger and colleagues connects these levels of evaluation. We propose bringing real work more fully into this approach through three contributions.
Identify how AI changes workload, autonomy, skills and working relationships. Involving the people affected helps compare promised gains with actual constraints and develop preventive measures before difficulties become established.
Document verification time, errors and situations in which decisions become difficult to challenge. These observations can inform interface design, limits on use and resources for oversight, and be discussed with project leaders and developers.
People who develop, test and evaluate AI need to be able to check results, report problems and challenge decisions. Their deadlines, resources and working conditions are a research direction for understanding the quality of safety checks.
The International AI Safety Report 2026 groups general-purpose AI risks into three overlapping families. The workplace examples below illustrate this framework.
Using AI to deceive or cause harm.
A fake voice message from a company director asks an employee to transfer money.
Errors and dangerous behaviours, including loss-of-control scenarios.
An agent changes case files without sufficient verification.
The effects of widespread AI adoption on the economy and society.
Task substitution may weaken skills and bargaining power.
The strength of evidence differs across risks. Observing a difficulty at work is not enough to attribute it to AI: the context, other changes and their evolution over time must also be examined.
An AI system that follows instructions is not necessarily safe. The approach combines evaluation, safeguards, monitoring and the ability to intervene, as described in the NIST framework. Different actors have complementary responsibilities.
Reduce dangerous behaviour, test system limits and document conditions of use. Provide mechanisms for control and correction, then use incidents and feedback from practice to inform improvements.
Define the need, assess effects on work and start with a limited trial. Give workers time and resources for oversight, establish ways to challenge decisions and decide when use should be changed or stopped.
Analyse work with employees and their representatives, monitor health effects and advise on adaptations. Feed relevant findings at the collective level back into the project while protecting the confidentiality of individual information.
Testing a model and assessing occupational risks are complementary activities. Strong technical performance does not establish that a use improves work; a prevention tool does not certify a model’s safety.
Links between work organisation and health are documented. Applying this knowledge to improve AI safety is a contribution that needs to be developed and evaluated. The framework proposed in Broutin and Faghri’s editorial is not a validated method; links between laboratory working conditions and model safety remain a research question.
Organisations to explore: INESIA · European AI Office · CeSIA.