Work intensity and time
AI may accelerate expected output, increase interruptions or add invisible verification work.
Thomas Cole, Destruction, 1836
IV. Risks & preventionAnticipate risks.
Anticipate risks.
Do not wait for harm to appear.
AI systems, work organisation and occupational health.
AI can change tasks, pace, discretion, cooperation and surveillance. Prevention should begin before deployment, with workers and occupational-health teams involved.
In 30 seconds
Essential points
- The central technical hazard is a plausible, fluent error that may escape routine checking.
- AI can affect all six major psychosocial-risk dimensions through pace, autonomy, relations, values, insecurity and emotional demands.
- Occupational health should be involved before deployment and contribute to monitoring effects on real work afterwards.
EstablishedEmergingAnalysis
On this page
The French edition includes the detailed mapping to the Gollac psychosocial-risk framework and complete references.
Psychosocial risks
How AI can act on psychosocial-risk pathways
The effect is not determined by the technology alone. It depends on implementation, staffing, objectives, performance metrics, participation and the resources available to workers.
Emotional demands
Customer-facing automation, content moderation or difficult interactions can redistribute emotionally demanding work toward humans.
Autonomy
Decision-support can broaden capability, but prescriptive systems may narrow discretion and make deviations harder to justify.
Social relations
Individualised metrics and reduced human contact can weaken cooperation, while poorly designed tools create conflict over errors and responsibility.
Ethical conflicts
Workers may be asked to endorse outputs they consider unsafe, unfair or inconsistent with professional standards.
Job insecurity
Restructuring narratives, skill obsolescence and uncertainty about future roles can become sustained stressors.
Technical risk
The plausible-error problem
A language model can produce a response that is grammatically confident, contextually plausible and factually wrong. This combination is hazardous because users may spend less attention checking a polished answer than an obviously poor one. Retrieval or web tools can reduce some errors but do not remove the need for verification.
Fabrication
The system may invent a source, rule, study, threshold or factual detail.
Contextual error
A proposition may be true in another country, profession or clinical situation but wrong for the case at hand.
Automation bias
Users may defer to a recommendation because it appears systematic or because challenging it is time-consuming.
Work system
Organisational, ethical and legal risks
Some harms are not experienced first as individual distress. They appear as poorer coordination, degraded service quality, data leakage, diluted responsibility or systematic disadvantage for particular groups.
Responsibility gaps
When several actors rely on a model, it may become unclear who was expected to verify, challenge or stop the decision.
Skill erosion
Routine delegation can reduce opportunities to practise, teach and maintain critical professional knowledge.
Surveillance
Monitoring tools may infer performance, behaviour or emotion from partial proxies and create continuous self-control.
Primary prevention
Prevention before deployment
Occupational health should be consulted early when a project may change tasks, pace, autonomy, monitoring, exposure to difficult interactions or employment prospects. The objective is not to certify the technology but to help analyse work effects and prevention measures.
Analyse real work
Observe current tasks, workarounds, interruptions, collective regulation and quality criteria before defining the future workflow.
Involve workers
Users and affected workers should be able to discuss the purpose, test conditions, error handling and reasons for contesting an output.
Protect verification time
Do not assume checking is cost-free. Allocate time, competence and access to evidence.
Set stop criteria
Define incidents or effects that pause the pilot, including data breaches, repeated critical errors, overload or unacceptable monitoring.
Follow-up
Monitor what changes after deployment
Post-deployment monitoring should combine technical incidents with work indicators: overrides, rework, verification time, workload, autonomy, conflicts, sick leave signals, perceived surveillance, unequal effects and loss of skill. Quantitative data should be discussed with qualitative observations.
Short-cycle review
Review the first weeks frequently, when workarounds and hidden burdens emerge.
Collective review
Share findings with worker representatives and prevention actors rather than limiting feedback to vendor or management dashboards.
Update risk assessment
Meaningful changes should feed the occupational risk assessment and prevention programme.
Key references
- Broutin C. et al., Large language models and new psychosocial issues at work.
- EU-OSHA resources on digitalisation, algorithmic management and occupational safety and health.
- Full references are available in the French reference edition.