AI safety work needs a psychosocial hazard register, not resilience messaging

The Financial Times reported on September 22 that some employees at AI companies and the UK AI Security Institute described stress, burnout or resignation linked to fears about advanced systems, rapid development and insufficient safeguards. The accounts span different organisations and roles; they do not establish prevalence or a single cause.

They do identify a management question that generic wellbeing programmes cannot answer. Safety researchers, red teamers, incident responders and policy staff may repeatedly encounter disturbing scenarios, ambiguous evidence and pressure to reach high-consequence judgements under time constraints. The hazard can arise from the work design, not from an individual's lack of resilience.

Register the exposure and the blocked route

Map tasks that involve sustained catastrophic-risk assessment, adversarial content, security incidents, moral conflict or responsibility without authority. For each task, record exposure duration, decision consequence, supervision, recovery time, escalation channel and whether the worker can pause work without penalty. Review teams and contractors as well as employees; outsourced exposure is still part of the operating model.

A psychological-safety survey alone is insufficient. Combine confidential pulse measures with workload, overtime, rotation, sick leave, attrition, unresolved escalation and retaliation complaints. Keep health data access tightly restricted and report only aggregates that cannot identify a small specialist team.

Protect dissent without turning it into evidence

Employees need a route to challenge a deployment or evaluation conclusion outside their reporting line. Record the claim, evidence requested, decision authority, response deadline and outcome. A protected challenge is not proof that the feared event will occur, and a rejected challenge is not proof that the concern was irrational. Preserve both the substantive safety analysis and the employment process.

Managers should distinguish three responses: immediate clinical or crisis support, temporary work adjustments, and a technical or governance review of the underlying concern. Conflating them can medicalise dissent or, conversely, leave a distressed employee carrying an unresolved system risk.

The countercase is that highly public debate about existential risk can amplify anxiety independent of workplace conditions. That is plausible and makes causal attribution difficult. It does not remove the employer's duty to assess controllable workload, role conflict, exposure and retaliation risk. Compare teams with different rotations, supervision and escalation designs rather than assuming a single narrative.

Pilot explicit exposure limits, paired review for high-consequence judgements, scheduled decompression and independent escalation. Evaluate error detection, unresolved concerns, absence, voluntary transfers and retention alongside wellbeing measures. Do not reward managers for suppressing reports.

This complements the principle that an AI safety protocol must be testable. Technical governance depends on people being able to surface weak signals without absorbing unlimited personal cost. A hazard register makes that dependency visible and gives leaders something concrete to change.

Give managers a decision protocol

When a worker reports distress and a technical concern together, the manager should acknowledge both, secure immediate safety, preserve relevant evidence and route each issue to the appropriate independent owner. Set maximum response times and prohibit performance penalties for good-faith escalation. Managers need training to avoid demanding repeated retellings of disturbing material or asking the affected employee to prove the entire systemic case alone.

Board reporting should show hazard exposure and control effectiveness without turning individual health into a governance metric. Report coverage of risk assessments, overdue actions, use of independent channels, rotation adherence and themes from anonymised cases. Include contractor populations and leavers where lawful. An external occupational-health review can test the control design, but it should not decide the technical validity of model-safety claims. Those two expert judgements must inform each other while remaining institutionally distinct.

Set a named owner and a review date for every proposed control. A recommendation without an accountable owner, evidence request and expiry becomes policy theatre. Preserve rejected alternatives and the reason for choosing the final design so later reviewers can distinguish a deliberate trade-off from an undocumented omission.