LLMs and psychosocial risks at work

The starting question

How can psychosocial risks be prevented when large language models transform intellectual tasks, occupations and professional relationships? The article examines the possible effects of their introduction into organisations and how occupational health professionals can support these changes.

The analytical framework

This narrative review brings together work on AI, research on working conditions and examples of organisational change. It draws on the six dimensions of the Gollac report: work intensity and working time, emotional demands, autonomy, social relationships, value conflicts and insecurity at work. The effects of LLMs are thus examined through risk factors already recognised in occupational health.

The central issue is organisational: automating a task alone does not determine its effect on health. The scope for professional judgement, production targets, training and workers’ participation influence how change is experienced.

Four risk mechanisms

1. Job insecurity and uncertainty about the future. Automating writing and analytical tasks may create a sense of being replaceable. Uncertainty about future responsibilities, expected skills or staffing levels can itself become a source of strain. The prospective studies cited describe potential task exposure; they should not be read as job losses that have already occurred.
2. Deskilling and loss of meaning. Delegating expert tasks over time may reduce opportunities to practise and maintain professional skills. The article connects this issue to autonomy, recognition and the ability to do work considered useful and of good quality. It draws in particular on LaborIA’s research into tensions between optimisation goals and what matters to professionals in their work.
3. Cognitive overload and stress. AI outputs must be checked, corrected and placed in context. This vigilance takes time and attention, while the tool’s apparent speed may lead to higher production demands. The article also discusses information overload, urgency and constant connectivity. Its comparison with alarm fatigue is an analogy intended to illuminate the work of supervision.
4. Isolation and deteriorating workplace relationships. Individual use of an assistant may replace previously collective exchanges. Automated monitoring, indicator-based assessment and unclear responsibility for errors may also undermine trust, support and cooperation. Maintaining opportunities for discussion and clear responsibilities is therefore essential.

Effects depend on conditions of use

The article emphasises the ambivalence of these technologies. Well-integrated AI can reduce repetitive tasks, free time for relationships, creativity or decision-making, and support skills. These benefits depend on training, participation in deployment and workers’ room for manoeuvre. Prevention therefore also involves identifying uses that strengthen people’s capacity to act.

The role of occupational health teams

Training should help professionals understand LLM capabilities and limitations, including biases and errors, then connect their use to real work. During consultations or workplace assessments, unusual workload, growing boredom, anxiety or sleep problems may prompt exploration of recent changes in activity, without automatically attributing these signs to AI.

The proposed prevention measures address work organisation: count verification time as part of workload, preserve exchanges between colleagues, clarify responsibilities and maintain opportunities to exercise expertise. This approach brings together occupational physicians, nurses, psychologists, ergonomists and other prevention professionals.

Six recommendations for organising prevention

1. Include the psychosocial effects of LLMs in the DUERP, France’s workplace risk assessment document, and link them to practical training and work-adaptation measures.
2. Develop coding for AI exposure in intercompany occupational health service databases to improve identification, traceability and collective analysis. This is a proposal made in the article.
3. Train all occupational health professionals, beyond physicians alone, and organise exchanges of practice.
4. Build tools for identifying risks: guides, questionnaires about workers’ experience, indicators and analysis of work after deployment.
5. Strengthen social dialogue on uses, workload, disconnection and maintaining skills.
6. Continue observation and research to follow effects over time and identify the sectors or populations most exposed.

Scope and limitations

This work offers a synthesis and directions for prevention. It does not present a worker cohort or an experiment measuring a causal effect of LLMs on health. Its references combine research, prospective reports and company examples with different levels of evidence. The effects discussed therefore need to be studied in each work context.

Summary of the full narrative review by C. Broutin, Archives des Maladies Professionnelles et de l’Environnement, 86 (2025), 102877. Access the publication ↗