AI for Psychosocial Hazards: Real Signal, or New Risk?

Yes, AI can read a quarter's worth of survey comments, absenteeism data and chat logs and tell you a team looks "at risk". I would still not let it decide who is struggling, or call that a control. The honest catch is this: under Australian WHS law the duty is to control the hazard, not to detect distress more efficiently. A monitoring tool bolted on without consultation doesn't reduce psychosocial risk. It can quietly add to it.

I'm a WHS professional, and managing psychosocial risk is now one of the most scrutinised duties in the country. The stakes are not abstract. Mental health conditions made up 12% of all serious workers' compensation claims in 2023-24, the highest proportion on record, and serious psychological injury claims have grown 161% over the past decade, the largest rise of any injury category (Safe Work Australia, 2025). So here is the field guide I'd actually use: where AI genuinely helps with psychosocial hazards, and where it becomes one.

What the law actually requires

Australian WHS law now treats psychosocial risk as a named, enforceable duty. Since the model WHS Regulations were amended to add psychosocial provisions (regulations 55A to 55D, commencing in the Commonwealth jurisdiction on 1 April 2023), a PCBU must identify psychosocial hazards and eliminate or minimise the risk so far as is reasonably practicable (Federal Register of Legislation, 2023). That duty is about the work, not the worker's mood, and it sits inside a hierarchy of controls that prioritises fixing the source.

The model code of practice, Managing psychosocial hazards at work, names the hazards a PCBU has to manage. They include:

• high or low job demands, low job control and poor support;
• lack of role clarity and poor organisational justice;
• bullying, harassment including sexual harassment, and violence or aggression;
• exposure to traumatic events or material.

Notice what they have in common. They are features of how work is designed and managed, not signals to be read off individuals.

Two complications matter before any tool enters the picture. First, WHS in Australia is a model-law scheme: the regulations have no force until each jurisdiction adopts them, so the exact wording and start date depend on where your workers are. NSW moved first, with an approved code in force from 28 May 2021 and the regulatory duty from 1 October 2022; Victoria, outside the harmonised scheme, was last, with standalone psychological-health regulations commencing 1 December 2025. As of mid-2026, every Australian jurisdiction has a psychosocial framework, but they are not identical.

Second, there's a separate duty people routinely conflate with this one. The Respect@Work positive duty under the Sex Discrimination Act 1984 (Cth) requires employers to take reasonable and proportionate measures to eliminate sexual harassment and sex discrimination. It commenced on 12 December 2022, and the Australian Human Rights Commission gained enforcement powers a year later, on 12 December 2023 (Law Society Journal, 2023). It overlaps with the WHS duty on harmful behaviour, but it has a different regulator and a different test. An AI tool can't tell the two apart. You have to.

Why psychosocial injury is the costliest claim category

Psychological injury is the most expensive, longest-duration claim category in the country, and it's still growing. A worker with a mental health condition claim was off work a median of 35.7 weeks in 2022-23, almost five times the 7.4-week median across all serious claims, and the median payout was $67,400 against $16,300 (Safe Work Australia, 2025).

The drivers are not mysterious, and they map cleanly onto the hazards the law names. Harassment and workplace bullying account for the largest share of serious mental-stress claims, followed by work pressure and exposure to violence.

What drives serious mental-stress claims in Australia

Category	Value (%)
Harassment & workplace bullying	33.2%
Work pressure	24.2%
Exposure to violence	15.7%

The cost lands well beyond compensation. In NSW alone, psychologically unsafe workplaces are estimated to cost $2.8 billion a year, and that's a conservative figure covering only claims, absenteeism and presenteeism (SafeWork NSW, 2024). Nationally, the Productivity Commission put lost workplace participation and productivity from mental ill health at up to $39 billion a year (Productivity Commission, 2020). When the problem is this large and this costly, the temptation to reach for a tool that promises to "see it coming" is obvious. That's exactly when judgement matters most.

Where AI genuinely helps with psychosocial hazards

AI earns its place as descriptive analytics on data you already hold, used to surface leading indicators at a team or system level. It's a faster way to triangulate signals a human reviewer would take longer to spot, never a verdict on a person.

Regulators already point to the raw material: absenteeism, turnover, grievances, complaints, workers' compensation data and consultation feedback are all named sources for identifying psychosocial hazards. The legitimate role for a model is to read across those at scale: cluster the recurring themes in free-text pulse-survey responses, flag a team where overtime has crept up for three months (a proxy for high job demands and low control), notice a turnover spike against a backdrop of negative exit-interview themes, or surface a cluster of grievances a single reviewer might read as one-offs. This is leading-indicator work, the same logic I set out in my piece on AI for safety analytics: measure the conditions upstream instead of counting injuries after the fact.

Three rules keep this honest. Work on aggregated, de-identified data with minimum group sizes, so you can't re-identify the person behind a number. Never score, rank or flag a named individual. And treat every output as a hypothesis to test through a human conversation, not a conclusion. Used this way, AI doesn't replace the risk-management process. It feeds it faster. The principle is the same one I apply across all of this work in the field guide to AI in workplace safety: AI belongs around the safety decision, not on it.

Can AI actually read how someone feels?

No, and this is the part vendors gloss over. There is no reliable way to infer a person's internal emotional state from their face, voice or word choice, and the science on this is not close. Emotion and sentiment AI measures patterns in expression and language. It does not measure feelings.

A landmark review of the evidence concluded that no facial-movement patterns map reliably, specifically and generalisably to emotion categories across people, cultures and contexts; the same scowl means many things (Barrett et al., Psychological Science in the Public Interest, 2019). When researchers tested automated facial emotion recognition across large, diverse samples, the task "crumbles into inconsistency", because even the human-labelled ground truth the systems learn from is unstable (Cabitza et al., Big Data & Society, 2022). Australians know it, too: only 12.9% support face-based emotion recognition at work (The Conversation, 2024).

Regulators are moving on it. From 2 February 2025 the EU AI Act prohibits AI systems that infer emotions in the workplace, except for medical or safety reasons, with penalties up to €35 million or 7% of global turnover (EU AI Act, Article 5, 2025). Australia has no equivalent ban yet. But a major jurisdiction has already outlawed the technique, and the peer-reviewed evidence says it doesn't work. That's a poor foundation for a psychosocial program.

When AI becomes the psychosocial hazard

AI becomes the hazard when monitoring intended to protect wellbeing turns into surveillance that erodes autonomy, control and trust. This isn't a hypothetical edge case. The Commonwealth code of practice, which commenced on 1 November 2024, now lists intrusive surveillance as a recognised psychosocial hazard in its own right, alongside low job control and poor support (Federal Register of Legislation, 2024).

Think about what continuous AI monitoring does to the conditions the law tells you to manage. It lowers job control. It can manufacture a chilling effect, where workers self-censor, stop raising issues, and quietly avoid the employee assistance program because they assume someone, or something, is watching. And it corrupts the very signal it relies on: if people know their survey comments and chat are being scored, they stop being candid, and the data degrades. The intervention can cause the harm it claims to measure. Australian workers feel this acutely.

Share of Australian workers concerned about intrusive monitoring

Category	Value (%)
Audio surveillance	72%
Live-screen monitoring	71%
Location tracking	71%
Emotion recognition	70%

Three in four surveyed workers say monitoring should only be used where it's clearly communicated and strictly necessary (Human Technology Institute, UTS, 2026). The regulatory direction is the same: the International Labour Organization's 2026 working paper names workplace surveillance, work intensification and reduced autonomy as AI-driven psychosocial risks (ILO, 2026), and NSW has passed first-of-its-kind laws on algorithmic management, which I covered in NSW's new AI work health and safety law. The trend is clear: regulators are increasingly treating monitoring as a hazard to manage, not a control to deploy.

Using AI here without losing rigour

Keep AI as an early-warning aid that feeds the risk-management process, and keep it strictly subordinate to the hierarchy of controls. The duty is to eliminate or minimise the hazard at its source. Detecting distress more efficiently, while leaving the workload, the bullying or the low control untouched, inverts the whole framework. An EAP, and an AI watching for who needs one, are support after harm, not a substitute for designing the harm out.

Here's the division of labour I'd actually run.

Step	What AI can do	What a human owns
Identify hazards	Cluster themes in aggregated survey, turnover and grievance data	Decide which signals are real, in consultation with workers
Assess risk	Surface trends and changes over time across teams	Judge severity, duration and who is exposed
Control	Nothing. Suggest nothing as "the control"	Fix work design, demands, support, behaviour
Individuals	Nothing. No scoring, ranking or flagging of named people	Have the conversation; offer support
Review	Track whether indicators move after a change	Decide if the control worked, and consult again

A few guardrails make the difference between an early-warning aid and a liability:

• Aggregate and de-identify. Team or cohort level only, with minimum group sizes. No individual scores, ever.
• Consult before you deploy. Rolling out a monitoring tool is itself a health-and-safety matter, so the duty to consult workers and HSRs applies. The consultation gap is exactly what erodes trust.
• Honour surveillance and privacy law. State surveillance regimes such as the NSW Workplace Surveillance Act 2005 require at least 14 days' written notice for overt monitoring, and inferred mental-health information is sensitive information under the Privacy Act 1988. The employee-records exemption is narrow and under reform; don't treat it as a safe harbour.
• Treat outputs as questions, not answers. Every flag is a prompt to investigate through a human conversation.

The bottom line

The problem is real and growing, and the data backs the urgency. The solution on offer is half-built and oversold. The credible role for AI in psychosocial work is narrow and genuinely useful: descriptive analytics on consented, aggregated data to spot leading indicators a human would take longer to see. The non-credible role, inferring individual emotions, predicting who will break, or passing off an algorithmic wellbeing score as a control, is bad science and worse safety practice. Leading indicators, not mind-reading.

This is the same lesson as my guide to managing AI itself as a WHS risk: the tool belongs around the decision, never on it. It echoes the test I apply to AI for incident investigation, where a confident wrong answer does the most damage, and it rests on the same foundation as grounding AI in the actual law. If you're working out where AI fits in your own psychosocial program, reach out. I'm always happy to compare notes.