AI and Psychosocial Hazards in Australian Workplaces: What Employers Must Do

Why AI creates psychosocial hazards

Psychosocial hazards are factors in work design, environment, and relationships that increase the risk of psychological harm. AI-driven work management creates several specific and documented psychosocial hazards that differ from those created by human management, primarily because of the scale, speed, and opacity of AI decision-making.

The academic literature on algorithmic management identifies the following as the primary mechanisms by which AI creates psychological harm in the workplace: surveillance — constant monitoring creates anxiety and a sense of being watched even when performing well; power asymmetry — algorithmic managers never take breaks, never make exceptions, and cannot be reasoned with in the way a human manager can; opacity — workers cannot understand or predict the algorithm's demands; pace intensity — algorithms optimise aggregate throughput without individual welfare consideration; and unpredictability — algorithmic scoring that fluctuates without clear cause creates anxiety and learned helplessness.

The regulatory framework: all jurisdictions now covered

As of 1 December 2025, all Australian jurisdictions have psychosocial hazard regulations. Victoria was the last, with the Occupational Health and Safety (Psychological Health) Regulations 2025 commencing on that date. The regulatory framework across jurisdictions requires PCBUs to: identify psychosocial hazards in their workplace; assess the risk of psychological harm they create; implement controls to eliminate or minimise those risks; and monitor and review the controls.

Victoria's regulations are notable for explicitly requiring employers to prioritise higher-order controls — changes to work design, systems, environment, and management — over lower-order controls like information and training. Telling workers to be resilient to algorithmic surveillance is not a compliant control measure in Victoria (or consistent with the hierarchy of controls elsewhere) if changes to the system itself would address the hazard.

AI-specific psychosocial hazards to assess

When conducting a psychosocial hazard risk assessment for workplaces using AI tools, specifically assess the following:

AI-driven productivity monitoring: Does the system measure individual worker output in real time? Does it create visible performance scores or rankings that create competitive pressure? Are targets set by the AI without human review of their achievability? Does the monitoring create a sense of constant surveillance?

Algorithmic scheduling: Does AI scheduling reduce worker control over their hours and roster? Does it create unpredictable shift patterns? Does it contact workers outside agreed hours to fill gaps? Does it optimise for business need without accounting for worker welfare?

AI performance management: Are performance ratings generated automatically without human interpretation? Are disciplinary processes triggered automatically by AI outputs? Can workers see and understand their AI-generated scores? Is there a genuine contestability process?

AI task allocation: Does the AI set unachievable workloads at certain times? Does it create a queue of demands that the worker cannot manage? Does it reduce worker discretion over how tasks are performed?

AI-generated feedback and communication: Does the organisation use AI to generate performance feedback? Do workers interact with AI systems that deliver bad news (performance warnings, shift cancellations) without human presence?

Controls: the hierarchy applied to AI hazards

Higher-order controls should be considered first:

Elimination and substitution: Can the feature that creates the hazard be removed or replaced? AI surveillance that measures every keystroke could be replaced with output-based performance metrics. AI scheduling that contacts workers at any hour could be configured with a contact curfew. If the hazardous feature is not essential to the business purpose, eliminate it.

Engineering controls: Build protections into the system. Mandatory rest alerts that the AI cannot override. Transparency dashboards that let workers see their AI-generated scores and the factors behind them. A built-in contestability workflow. Volume limits on AI-generated task queues per worker per hour. These are engineering solutions that reduce the hazard at the design level.

Administrative controls: Establish clear policies about AI monitoring — what is measured, how it is used, and what it cannot trigger automatically. Require human review of AI-generated performance data before it is used in disciplinary processes. Provide training for workers on how the AI system works and how to raise concerns. Create formal mechanisms for workers to flag unsustainable AI-generated workloads.

Training and information: These are the lowest level of control and should not be the primary response. Training workers to be psychologically resilient to AI surveillance is not an adequate control measure. However, it is a valuable complement to higher-order controls.

Engaging workers in the assessment

Workers experience psychosocial hazards that managers and HR cannot see from outside the workflow. A productivity target that looks reasonable from the management dashboard may feel like a relentless pressure from inside the role. Include workers and HSRs in the hazard identification process — they will identify risks that the hazard assessment would miss.

The psychosocial regulations in most jurisdictions require consultation with workers and HSRs on hazard identification and control. This is not a formality — it is the primary mechanism by which AI-driven workplace hazards will be identified and appropriately controlled.