When the System Becomes the Hazard: AI and Psychosocial Safety at Work
From the outside looking in
As a mediator and conciliator with over a decade of experience resolving community, workplace and commercial disputes, I spend much of my time on the outside of other people's workplaces, usually after a conflict has already escalated. That outside view is useful, because it lets me see patterns that are hard to notice from the inside. It has also taught me that two ideas leaders often treat as the same thing are in fact quite different, and that the difference now matters legally as well as culturally.
The first is psychological safety, the sense that a person can speak up, ask a question or admit a mistake without being humiliated. The second is psychosocial safety, which in Australia is a work health and safety obligation, with codes and regulations behind it. Artificial intelligence sits right where these two meet. Used without enough care, it can quietly become a psychosocial hazard in its own right. In this article I set out the difference, what the law now requires, and where AI fits in.
Psychological safety and psychosocial safety are not the same thing
It is worth separating these clearly, because they are managed in very different ways.
Psychological safety describes how a team feels. It is about whether people believe it is safe to disagree, raise a concern or admit that they are struggling. It is built slowly, through the way leaders respond in everyday moments, and it can be damaged quickly. It matters a great deal, but on its own it is not a legal standard.
Psychosocial safety is different. It is about how a workplace identifies and controls psychosocial hazards, the aspects of work design, work management, the working environment and workplace interactions that can cause psychological or physical harm. In Australia, managing these hazards is a legal duty, not an optional extra.
The practical point is a simple one. A team can feel reasonably comfortable day to day and still be exposed to psychosocial hazards through the way the work itself is set up. A positive culture is welcome, but the law asks about the structure of the work, not only the mood in the room.
What does Australian law actually require?
Under our work health and safety laws, a person conducting a business or undertaking, in most cases the employer, must eliminate psychosocial risks so far as is reasonably practicable, and where that is not possible, minimise them so far as is reasonably practicable. This was made explicit when the WHS Regulations were amended in 2022.
Here in Queensland, where I am based, the Managing the Risk of Psychosocial Hazards at Work Code of Practice 2022 and its regulation commenced on 1 April 2023. The other states and territories have moved in the same direction, and by the end of 2025 psychosocial regulations were in force across the country. The Commonwealth code goes a step further again, naming fatigue, intrusive surveillance and job insecurity as examples of hazards.
Two related duties sit alongside this framework. The positive duty under the Sex Discrimination Act 1984 (Cth) requires organisations to take reasonable and proportionate steps to eliminate sexual harassment and related conduct, and the Australian Human Rights Commission has had enforcement powers since December 2023. Internationally, ISO 45003 offers detailed guidance on managing psychosocial risk that lines up closely with the Australian codes.
What all of these have in common is that they are risk management duties. They ask you to identify the hazards, assess the risk, put controls in place, and review whether those controls are actually working. They are not satisfied by a values statement or a single training session.
Where does AI come into this?
When most people picture a psychosocial hazard, they think of bullying, harassment or a traumatic event. Those matter. But the codes are much broader, and they include quieter risks such as high or low job demands, low job control, poor support, unclear roles, poorly managed change, weak recognition and a lack of procedural fairness. This is exactly the territory where AI and automated systems are now reshaping the way we work.
In February 2026, New South Wales passed the Work Health and Safety Amendment (Digital Work Systems) Act 2026, the first Australian law to name algorithms, artificial intelligence, automation and online platforms as sources of work health and safety risk. It makes clear that employers must ensure their use of these systems does not expose workers to harm, including through workloads set by algorithms, an over-reliance on performance metrics, and decisions made unfairly by biased systems. Importantly, that duty stays with the employer even when the tool has been supplied by a third party. You cannot contract the risk out to a vendor.
The distinction I would draw is that the hazard is rarely the technology itself. It is the way the technology is built into the work. The same scheduling tool can lighten a workload or make it relentless. The same dashboard can support a worker or keep watch over them. A few common examples show how easily ordinary practices can tip over into recognised psychosocial hazards:
Algorithmic work allocation and scheduling: unreasonable or unpredictable job demands, and low job control.
Continuous productivity and location monitoring: intrusive surveillance, chronic stress and low trust.
Automated performance metrics and ranking: excessive performance pressure, and low reward and recognition.
Automated or AI-assisted decisions about people: poor organisational justice, perceived unfairness and bias.
Rapid AI-driven change to roles and workflows: poorly managed change, job insecurity and fatigue.
Why good intentions are not enough
In my experience, most of the harm connected to AI will not come from bad intent. It will come from systems introduced to improve efficiency, with the effect on people left unexamined. Regulators are pointing to the same thing. SafeWork NSW guidance highlights design flaws such as machine-paced work with no room for recovery, systems that remove a worker's judgement, and automated feedback delivered with no regard for a person's circumstances. In one 2025 case, SafeWork NSW paused a university restructure because of the risk of psychological harm from inadequate consultation.
There is a lesson here that is older than AI. Change is difficult, particularly when people have no control over it and cannot see the reason for it. When a new system arrives without explanation, and feels as though it has been done to people rather than with them, frustration turns into resistance and mistrust. Automation can speed that up, because decisions feel faster, less personal and harder to question.
This is also why written policies, on their own, will not be enough. The duty is to change the way the work is actually done, not simply to describe good intentions in a document.
Keeping the person and the system in view
A useful question, when conflict or distress shows up around a new system, is not only whether someone is behaving badly. It is whether the work itself has become harder to do well. People want some say in how they work, and they want to feel that they are good at it. When an automated system removes both, something important is lost, and it often shows up as friction between people long before anyone points to the system as the cause.
Culture matters here too. It is built slowly and can be damaged quickly. A monitoring tool that signals distrust, or an automated decision that feels unfair, can undo months of careful relationship building in a single rollout. Looking only at the individuals in conflict, and not at the system they are trying to work within, usually misses the real hazard.
Two scenarios from practice
The following are de-identified, composite scenarios. They do not describe any single organisation or person.
Scenario one: reading what the machine wrote. A complaints and review team notices that the correspondence coming in has changed. More of it is drafted or expanded using AI, so submissions are longer, more repetitive and harder to follow. The case managers are still expected to read every word carefully, because procedural fairness requires it, but they have been given no guidance on how to manage the volume, how much weight to give material that may be machine-generated, or whether to raise the use of AI at all. The work has quietly grown, the emotional load has grown with it, and staff are tired in a way that is hard to explain to anyone simply counting closed files.
Looked at through the codes, the hazard here is not the technology. It is high job demands, unclear expectations and inadequate support, left unmanaged. Clear guidance, a sensible way to triage, and a stated position on AI-assisted correspondence are not administrative niceties. They are control measures.
Scenario two: permission to innovate, with no way through. In another organisation, staff are encouraged to find responsible ways to use AI. Some take that seriously, put real effort into identifying safe and useful applications, and submit them through the approval process. What follows is a long governance pathway, then repeated requests for more explanation, more assurance and more paperwork, and no decision. The proposals are not refused. They simply never clear the queue. Over time, the people who put themselves forward learn that showing initiative creates work and exposure for them and little else, so they stop.
Here the harm sits in low job control, weak reward and recognition, and poor organisational justice, the sense that effort and fairness have come apart. There is a quieter cost as well. A process that invites people to contribute and then never resolves anything slowly wears down their willingness to try. A governance process is appropriate and necessary. A governance process that only ever creates a bottleneck becomes a hazard of its own.
Compassion fatigue, and the risk of AI talking to AI
A good deal of my training and supervision work is with mediators, conciliators and complaints professionals, and compassion fatigue comes up again and again. In conflict work, some emotional wear is close to unavoidable. You cannot sit with other people's distress, week after week, and remain completely untouched by it. One of the strongest protections against that fatigue is meaning, the sense that the work matters, that a real person was heard, and that something was understood or repaired.
This is where I have a real concern about AI in dispute resolution. When correspondence is generated by AI on one side, and sorted or answered by AI on the other, the human contact that gives the work its meaning can quietly disappear. If a practitioner spends the day responding to machine-generated volume and never quite reaches the person underneath it, the work becomes harder to engage with and harder to value. That is not only a wellbeing concern for the individual. It is a quality concern for the organisation, because tired and disengaged people add less value at exactly the moments that call for careful human judgement.
The point is not that AI has no place in this work. It is that we need to design these processes so that people can still see the human at the other end, and still feel that their contribution counts. Protecting meaning is part of protecting the workforce, and it belongs in the same conversation as workload and surveillance when we think about psychosocial risk.
What can leaders do now?
The codes point to a straightforward, four-step process, and it works just as well for AI as it does for any other hazard.
Identify. Map the digital systems that allocate work, set the pace, monitor people, or make or support decisions about them. Name the psychosocial hazards each one could create.
Assess. Weigh up how likely and how serious the harm could be, and look at the hazards together rather than one at a time. Talk to the people who actually use the system, because they notice the pressure points first.
Control. Change the design of the work, not just the language around it. That might mean adjusting workloads, building in time to recover, keeping human judgement in decisions, and being open about what is monitored and why.
Review. Check whether the controls are working, and adjust them where they are not. Keep a simple record of what you considered and did, so you can show that you took reasonable steps.
None of this needs a large budget or an outside consultant to get started. What it needs is genuine curiosity about how people are experiencing the work, and a willingness to change the parts that are causing harm. In conflict work, clarity is part of care.
Where this leaves us
Psychological safety and psychosocial safety are related, but they are not the same. One is about how a team feels. The other is about how work is designed and managed, and it now carries clear legal duties in Australia.
The encouraging news is that most of this is workable. When AI is treated as a change to the system of work, named early and managed with structure rather than blame, it can be introduced in a way that protects both wellbeing and performance. Much like a workplace dispute, the earlier these risks are addressed, the less time, money and goodwill is lost along the way. If your organisation is bringing in new tools and you would like to think through the people and process risks before they become disputes, I would be glad to have a confidential conversation about the options.
Sources and further reading
This article is general information, not legal advice. Work health and safety duties differ between jurisdictions and change over time, so seek advice on your specific circumstances.
.png)