From the model to the work system
AI safety commonly studies model reliability, robustness, control and misuse. At work, these properties describe only part of the risk. AI enters an organisation shaped by targets, deadlines, power asymmetries and rules of responsibility.
The system actually deployed is the interaction between model, task, workers, organisation and implementation choices. Its safety cannot be inferred from a technical benchmark.
Health as a property of deployment
Cognitive load, intensification, surveillance, autonomy, prevented quality and employment insecurity can change without a spectacular incident. Yet these gradual effects are safety failures when they durably damage health or the capacity to act.
Extending AI safety to occupational health means observing these transformations before deployment and then in real activity, with workers and prevention professionals.