AI becomes a psychosocial risk when it disrupts work organisation.
Written by Dr Charles Broutin, occupational physician, AI lead at the French Society of Occupational Health · Updated on
By redistributing tasks, deadlines, control, responsibility, relationships and skills, an artificial intelligence (AI) system can create new demands or weaken the resources available to cope with them. The risk therefore comes not from the model alone, but from how work is organised around it.
Scientific sourcesGollac frameworkThree levels of preventionCaution about causality
Direct answer
The essentials
Artificial intelligence rarely creates an entirely new risk: it reshapes known psychosocial risk factors when it brings lasting changes to work organisation, whether through workload, autonomy, responsibilities, relationships, skills or job security. The available studies do not show a single, uniform effect: some uses improve productivity or control over work, while others are associated with greater surveillance, invisible work, loneliness or exhaustion. Prevention must therefore examine the whole work system, with the people concerned, using the six dimensions of the Gollac report, and then monitor actual effects over time.
01
The risk is not in the model alone. It arises from the interaction between a technical capability, a task, targets, people and an organisation.
02
Findings vary. The cited studies include 5,172 support agents, 197 Australian workers and randomised trials in healthcare.
03
Prevention remains collective. It describes work before and after deployment, monitors warning signs across the six Gollac dimensions, and acts at three levels: at the source, through early detection, and through support.
01 / Six Gollac dimensions
How does AI change the six psychosocial risk factors?
An AI tool rarely creates an entirely new risk: it reshapes known factors. The six dimensions of the Gollac report describe work organisation and employment, not workers’ presumed vulnerability, and a single deployment can change several dimensions at once.
An appointment with the occupational physician on request
Support for returning to work after exhaustion
References: French Labour Code, L. 4121-2 (principles of prevention) · L. 2242-17 (right to disconnect) · AI Act, Article 4 (AI literacy).
02
Emotional demands
Contact with the public and with suffering, the requirement to control or hide emotions, and fear at work.
Diagram · Who receives which requests?
Volume falls, but the share of difficult interactions handled by people grows. Illustrative diagram.
What AI changes
Concentration of difficult cases
AI handles simple cases; people receive the contentious and emotionally charged ones.
Cascading frustration
The user arrives annoyed after the chatbot has failed.
Justifying the unjustifiable
Defending an inexplicable algorithmic decision to the public.
Emotions under surveillance
Tone analysis and automated scoring of interactions.
Fear
Of being held responsible for the tool’s errors, or of losing professional standing.
Warning signs
Increasing incivility
Exhaustion and cynicism
Requests for transfers
More frequent tense interactions
PrimaryAct at the source
Allocate work between people and AI according to emotional demands
Ensure quick and easy access to a human contact
Do not use emotion recognition on employees
SecondaryIdentify and strengthen
Alternate high-exposure tasks with less demanding ones
Provide training for tense situations
Set up groups for reflecting on professional practice
TertiaryProvide support
Debriefing and support after an incident
Follow-up by occupational health and prevention services, with psychological referral if needed
References: AI Act, Article 5 (prohibition of emotion recognition in the workplace) · French Labour Code, L. 4121-1 · GDPR, Article 22.
03
Autonomy
Discretion, participation in decisions, and opportunities to use and develop skills.
Diagram · Who decides? The automation ladder
The higher AI moves up the decision ladder, the more human oversight needs to be properly supported and effective. Illustrative diagram.
What AI changes
Algorithmic management
Tasks assigned, sequenced and evaluated by a system.
More prescriptive work
The suggestion becomes the norm; departing from it requires justification.
Automation bias
It is difficult to challenge a machine perceived as reliable.
Loss of skills
Skills that are no longer practised erode, and dependence develops.
Detailed tracking
Recording every action reduces the space for initiative.
Warning signs
Feeling that one carries out tasks without making decisions
Recommendations followed automatically
Work blocked by system failures
Loss of interest in the profession
PrimaryAct at the source
Involve workers in choosing and configuring the tool
Recognise the right to reject a recommendation without penalty
Ensure effective human oversight: time, competence and authority
SecondaryIdentify and strengthen
Provide training on AI’s limitations and typical errors
Value justified departures from recommendations
Organise regular reviews of experience
TertiaryProvide support
Support workers experiencing a loss of skills or meaning
Skills assessments and training leading to qualifications
References: AI Act, Articles 4, 14 and 26 (AI literacy, human oversight, deployer obligations) · GDPR, Article 22 · French Labour Code, L. 2312-8.
04
Social relationships at work
Relationships with colleagues and managers, support, recognition, fairness and management quality.
Diagram · The team’s network, before and after
When mutual support goes through the tool, direct relationships and the transmission of skills weaken. Illustrative diagram.
What AI changes
Algorithmic evaluation
Opaque indicators perceived as unfair or biased.
Isolation
People ask AI rather than a colleague.
Recognition diverted
The result is credited to the tool; human work becomes invisible.
Divisions
Advanced users versus non-users, and hidden uses.
Remote management
Management through dashboards, removed from actual work.
Warning signs
Disputes over evaluations
Fewer informal exchanges
Undeclared AI use
Complaints about unfairness
PrimaryAct at the source
No entirely algorithmic individual evaluations
Criteria that can be explained and challenged
An AI use policy that provides a framework rather than a ban
SecondaryIdentify and strengthen
Create spaces to discuss work
Train managers to recognise actual work
Appoint local AI contacts and organise mentoring
TertiaryProvide support
Mediation in conflicts
Care for work-related distress
References: GDPR, Article 22 (automated decisions) · AI Act, Article 26 · French Labour Code, L. 1132-1 (non-discrimination), L. 2312-8.
05
Value conflicts
Ethical conflicts, barriers to quality work and a sense of doing pointless work.
Diagram · The chain of responsibility
The worker who approves bears responsibility for the decision without controlling the preceding stages. Illustrative diagram.
What AI changes
Barriers to quality work
Producing quickly, without being able to do what one considers good work.
Contested decisions
Applying sorting, rejections or priorities considered unfair.
Unclear responsibility
Who is accountable for an AI error that someone has approved?
Sensitive data
Information entrusted to a tool without sufficient safeguards.
Loss of meaning
Work reduced to checking the machine, or even preparing one’s own replacement.
Warning signs
Talk of “rushed, poor-quality work”
Workarounds that bypass the tool
Unspoken ethical disagreements
Disengagement and departures
PrimaryAct at the source
Define quality in AI-assisted work collectively
Document the allocation of responsibilities
Make it possible to flag and reject a recommendation
SecondaryIdentify and strengthen
Organise discussions about the quality of work
Appoint an ethics contact who involves the professions concerned
Provide training on model biases
TertiaryProvide support
Support workers experiencing ethical distress
Medical follow-up and referral if needed
References: GDPR, Article 35 (impact assessment) · AI Act, Article 26 · French Labour Code, L. 4121-2.
06
Job and work situation insecurity
Socio-economic insecurity (employment, pay and career) and poorly prepared or poorly managed changes in work.
Diagram · The change curve
A well-prepared transformation causes a smaller loss of trust and restores it faster. Illustrative curve, not measured data.
What AI changes
Fear of replacement
Fuelled by public discourse and uncertainty about internal projects.
Rapid changes
Tools deployed or replaced without a clear timetable.
Perceived obsolescence
The feeling that one’s skills are losing value.
Vague announcements
A “transformation” with no clear account of its effects on roles.
Technical dependence
Outages, a change of supplier, or loss of control over the tool.
Warning signs
Rumours and concern
Repeated questions to the CSE
Anxiety and sleep problems
Early departures
PrimaryAct at the source
Provide clear information early and consult the CSE
Include AI in workforce and career planning (GEPP): career paths and retraining
Deploy in stages, with the option to reverse course
SecondaryIdentify and strengthen
Make training available to everyone, including less-qualified workers
Use professional development interviews to discuss changes in the role
Monitor workforce indicators
TertiaryProvide support
Support workers whose roles are changing
Prevent loss of employment due to health problems with occupational health and prevention services
References: French Labour Code, L. 2312-8 (CSE) · L. 2242-20 (workforce and career planning, GEPP) · L. 6315-1 (professional development interviews) · AI Act, Article 26.
Tool · Assess deployment
Assess your AI project through actual work.
A collective prevention tool focused on work organisation, never on individuals: nine dimensions, two questions per dimension, around ten minutes. Complete it before the pilot (T0), then over time (T1, T2+). Responses stay in the browser, and the report can inform updates to the occupational risk assessment document.
AI can accelerate production, impose a recommendation, monitor activity or automate part of a judgement. If the organisation raises targets, reduces discretion, makes checking work invisible or sustains uncertainty, this transformation can contribute to psychosocial risk factors (PSRs). To relate this analysis to a concrete situation, see the pathway on preventing AI-related risks.
01 / SYSTEM
A capability is introduced.
Generating, sorting, recommending, monitoring or contributing to a decision.
02 / ORGANISATION
Work is redistributed.
Tasks, targets, deadlines, responsibilities, control, cooperation and learning change.
03 / EXPOSURE
Demands and resources change.
Intensity, autonomy, support, value conflicts, insecurity and invisible work are reshaped.
04 / HEALTH
Effects may emerge.
Depending on duration, accumulated exposure and scope for action: stress, exhaustion, sleep problems or cardiovascular harm.
02 / Understanding the scale
Epidemiological figures to read without attributing them to AI.
These figures show the health significance of psychosocial working conditions. Apart from the occupational exposure indicator, they do not measure an effect caused by artificial intelligence.
According to the ILO, in 2025, one in four workers worldwide has an occupation with some degree of potential exposure to generative AI; the organisation considers task transformation more likely than widespread replacement. Read the ILO source ↗
1 IN 4
Jobs potentially exposed.
Worldwide, one in four workers has an occupation with some degree of potential exposure to generative AI.
ILO, 2025 · Task transformation is considered more likely than widespread replacement.> 840,000
Deaths associated with psychosocial risks.
The ILO estimates that psychosocial risk factors at work are linked to more than 840,000 deaths a year and nearly 45 million healthy life years lost.
ILO, 2026 · Global estimate across all psychosocial exposures.12 BILLION
Working days lost.
Depression and anxiety, from all causes, are estimated to account for 12 billion lost working days each year.
WHO, 2024 · Estimated cost: US$1 trillion in lost productivity.745,000
Deaths linked to long working hours.
WHO and the ILO attributed 745,000 deaths from ischaemic heart disease or stroke to exposure to long working hours in 2016.
WHO–ILO · 23.3 million DALYs; a reference point on work intensity, not on AI.
An essential caution: these figures cannot be used to estimate an “AI burden of disease”. They explain why a technology that brings lasting changes to workload, autonomy, working hours or insecurity must be studied as a transformation of working conditions.
03 / Findings specific to AI
What effects of AI is research already observing?
Studies examine different tools, occupations, time periods and outcomes. They describe possibilities, not a universal consequence of AI use.
In the study Generative AI at Work by Brynjolfsson, Li and Raymond, a generative AI assistant increased the number of issues resolved per hour among 5,172 support agents, with particularly large gains among less-experienced workers. Read the study ↗
PRODUCTIVITY
Real gains in some tasks.
Experimental writing tasks were completed faster, with improved average quality.
Among 5,172 support agents, an assistant increased the number of issues resolved per hour.
Gains were particularly large among less-experienced workers.
FAVOURABLE EFFECTS
Control and exhaustion can improve.
Among 197 Australian workers, generative AI use was associated with greater control over work, with no significant association with burnout.
In a randomised healthcare trial, ambient AI reduced exhaustion, interpersonal disengagement and daily time spent writing notes.
Another trial reduced documentation time for one of two systems; findings on burnout remain exploratory.
ADVERSE EFFECTS
Other studies identify vulnerabilities.
Algorithmic management is associated with perceived threat and burnout in a cross-sectional study.
Passively copying generated content reduced self-efficacy, a sense of ownership and task meaning in a preregistered experiment.
Greater interaction with AI was associated with loneliness and adverse after-work effects in four studies.
Time saved is not yet a health benefit. It becomes one if the organisation reinvests it in recovery, learning, quality, relationships or sustainable staffing. Immediately converted into higher targets, it can instead fuel work intensification.
04 / Prevention
Act at three levels, before, during and after deployment.
Preventing AI-related psychosocial risks follows the logic of any transformation of work: act first on work organisation and the choice of tool, identify deterioration early, then support people already affected.
Primary prevention
Act at the source.
Address work organisation and the choice of tool before exposure becomes established.
Allocate tasks between people and AI according to actual workload and emotional demands
Decide collectively how to use the time saved
Recognise the right to reject a recommendation
Train during working hours
Secondary prevention
Identify and strengthen.
Detect warning signs early and strengthen the resources workers have to cope with them.
Monitor workload and workforce indicators
Create spaces to discuss actual work
Train managers to recognise actual work and overload
Appoint local AI contacts
Tertiary prevention
Provide support.
Care for people already affected and organise their continued employment or return to work.
An appointment with the occupational physician on request
Debriefing and support after an incident
Support for returning to work after exhaustion
Preventing health-related loss of employment
Primary prevention comes first. The French Labour Code requires avoiding risks, tackling them at their source and adapting work to the individual (Article L. 4121-2). Training workers to manage stress does not compensate for an organisation that exposes them to it.
Step 1
Before deployment
Analyse the actual work involved and place the tool on the automation ladder. Involve workers and consult the CSE on introducing the new technology.
Step 2
During use
Monitor warning signs across all six dimensions and discuss them collectively. Measure actual workload, including checking time.
Step 3
Afterwards, and whenever something changes
Record identified risks in the occupational risk assessment document and reassess them with each update or change of tool. Keep deployment reversible.
05 / Reference framework
The legislation underpinning prevention.
Three bodies of law complement one another: the French Labour Code sets prevention duties and the role of the CSE, the European AI Act governs uses, and the GDPR protects people in relation to automated decisions. Details are available in the guide Law and governance of AI at work.
French Labour Code
L. 4121-1 and L. 4121-2
The employer’s health and safety duty and the general principles of prevention
R. 4121-1 et seq.
The occupational risk assessment document (DUERP)
L. 2312-8
Consultation with the Social and Economic Committee (CSE) on introducing new technologies
L. 2242-17
The right to disconnect
L. 2242-20
Workforce and career planning
L. 6315-1
Professional development interviews
L. 1132-1
The principle of non-discrimination
Regulation (EU) 2024/1689 on AI
Article 4
AI literacy for people using the systems
Article 5
Prohibition of emotion recognition in the workplace, except for medical or safety reasons
Article 14
Human oversight of high-risk systems
Article 26
Deployer obligations, including informing worker representatives and the workers concerned
GDPR
Article 22
The right not to be subject to a decision based solely on automated processing
Article 35
Data protection impact assessment
06 / State of the evidence
Prevent harm without claiming that everything has already been proven.
Research on the mental health effects of specific deployments remains limited, fragmented and heterogeneous. Causal inferences are constrained by the methods, contexts and outcomes studied.
ESTABLISHED
Working conditions matter for health.
High demands combined with low control are linked to clinical depression and coronary heart disease.
Insecurity, organisational injustice and low support have documented associations with health outcomes.
Prevention must target organisational factors, not only individual adaptation.
EMERGING
AI redistributes several exposures.
Checking, monitoring and responsibility can create extra work.
Benefits, risks and the groups affected vary considerably across work systems.
Favourable and adverse effects can coexist within the same organisation.
UNKNOWN
The health impact specific to AI.
No data allow a global burden of disease attributable to AI at work to be extrapolated.
Long-term effects and causal relationships require longitudinal studies and intervention trials.
Independent access to deployment data remains necessary.
Limited evidence is neither proof of inevitable harm nor proof of no risk. It calls for proportionate prevention, observation over time and the ability to revise the work system. The diagrams on this page are illustrative: they describe plausible mechanisms, not measurements.
07 / Scientific sources
Find the evidence behind the claims.
The references below lead back to the institutional reports, scientific studies and legislation on which this synthesis is based.
Document the task, the authority given to the system, the expected effects and the conditions required to deploy, change course or stop. Explore the AI project assessment method →