NSW's New AI Work Health and Safety Law, Explained

In February 2026 New South Wales did something no other Australian jurisdiction has done: it wrote artificial intelligence directly into work health and safety law. The Work Health and Safety Amendment (Digital Work Systems) Act 2026 (NSW), Act No 5 of 2026, received assent on 18 February 2026. It makes the safety risks of algorithms, AI, automation and online platforms an explicit part of a PCBU's duty of care (NSW Parliament, February 2026). It is not yet in force, and that gap is the point of this post: there is still time to get ready before the duty switches on.

I've spent a decade in WHS, across enterprise construction and social infrastructure, and I build AI tools grounded in the legislation itself, including a WHS professional skill encoded from the Acts. So I read this one twice: once as a practitioner who will have to advise on it, and once as someone who builds the kind of system it now regulates. Here is what it says, what it does not, and what I'd do about it now.

What did NSW actually pass?

NSW amended its existing Work Health and Safety Act 2011, rather than writing a standalone AI statute. The Digital Work Systems Act inserts new duties into the WHS Act and defines its key term broadly: a digital work system means "an algorithm, artificial intelligence, automation or online platform" (NSW Parliament, Digital Work Systems Bill 2026, Second Print, 2026). Those four words are doing a lot of work.

The Bill was introduced on 19 November 2025 by Sophie Cotsis, agreed by the Legislative Council on 12 February 2026, and assented on 18 February as Act No 5 of 2026. The definition is deliberately wide. It catches the warehouse rostering algorithm, the productivity-scoring dashboard, the delivery-platform dispatch engine and the tool that screens job applicants, not just the systems your organisation has badged as "AI".

Is the new duty in force yet?

No. As of June 2026 the Act has been passed and assented, but it is not yet operative. Section 2 states the Act "commences on a day or days to be appointed by proclamation", and no proclamation date has been announced (NSW Parliament, Second Print, 2026). The substantive duties are simply not switched on.

The commencement is also staged. Legal analysis of the Act notes that the new workplace-entry powers, which let union officials inspect a digital work system, do not start until at least a month after SafeWork NSW publishes guidelines on those powers, and those guidelines follow public consultation that has not happened yet (Norton Rose Fulbright, 2026). One firm put it plainly for employers: "none of the provisions in the AI Bill are in effect yet" (Colin Biggers & Paisley, February 2026).

So no one is in breach today. But "not in force" is not the same as "not your problem". The lead time before proclamation is precisely when a careful PCBU does the unglamorous work, and the assessment is the slow part, not the compliance date.

What does the new duty actually require?

The Act adds the obligation in two places. First, it widens the primary duty of care (a new section 19(3)(c1)) so a PCBU must ensure, so far as is reasonably practicable, that workers' health and safety is not put at risk from the use of digital work systems. Second, a new section 21A creates a specific duty about how a digital work system allocates work, and it names the risks a PCBU has to turn its mind to (NSW Parliament, Second Print, 2026):

• excessive or unreasonable workloads
• excessive or unreasonable performance metrics or targets
• excessive monitoring or surveillance
• unlawful discriminatory practices or decision-making

Read that list again and notice what it is. This is not a blanket "AI safety law" trying to govern every conceivable model risk. It is squarely aimed at how people are managed: the pace the system sets, the targets it enforces, the surveillance it runs, and the fairness of the decisions it makes about workers. That is psychosocial territory, and it is no accident.

The Act's risk category	What it looks like on the ground
Excessive or unreasonable workloads	A dispatch algorithm packing jobs so tightly there is no recovery time
Excessive or unreasonable performance metrics	A scoring system that ranks and punishes against targets a human never sanity-checked
Excessive monitoring or surveillance	Keystroke, location or productivity tracking that runs well past what the task needs
Unlawful discriminatory decision-making	An automated screen or allocation that quietly disadvantages a protected group

Why does this matter before it even commences?

Because algorithmic management is already everywhere your workers are. In late 2025 the OECD found that 90% of US firms had adopted at least one tool to instruct, monitor or evaluate workers, against a 79% average across surveyed European countries and 40% in Japan. Nearly two-thirds of managers using these tools reported at least one concern (OECD, How widespread is algorithmic management in workplaces?, December 2025).

Firms using at least one algorithmic management tool, 2025

Category	Value (%)
United States	90%
Europe (average)	79%
Japan	40%

These systems are already in NSW workplaces, and their risks already map onto hazards NSW regulates. Since 2022 the model WHS Regulations and the NSW psychosocial code have required PCBUs to manage psychosocial hazards: workload, low job control, monitoring, unreasonable demands. The Digital Work Systems Act does not invent those risks. It names the technology that now amplifies them, and it makes the link explicit rather than implied.

The system it lands in is already strained. Mental health conditions made up 10.5% of serious workers' compensation claims in 2022-23, about 14,600 claims, a rise of 97.3% over the decade, with median time lost more than five times that of physical injuries (Safe Work Australia, Key Work Health and Safety Statistics 2024, 2025). A law that asks employers to watch the workload and surveillance their software imposes is not regulating a hypothetical.

There is a human pattern underneath it too. In KPMG and the University of Melbourne's 2025 study, 57% of Australian employees said they had relied on AI output at work without checking its accuracy, and 59% said they had made mistakes because of it (KPMG and University of Melbourne, Trust, attitudes and use of AI: A global study 2025, 2025). The law is catching up to a behaviour that is already common: trusting the system without checking what it does to people.

How should PCBUs prepare?

Start now, because the slow part is the assessment, not the compliance date. Five steps I'd run, in order:

1. Inventory your digital work systems. Rostering, productivity scoring, monitoring and surveillance, automated screening, dispatch and platform allocation. Given that broad definition, most organisations have more of these than they think.
2. Assess each against the four risk categories. Where does the system set the pace, the target, the surveillance level or the allocation, and what does that do to the people on the receiving end?
3. Consult the workers who do the work. Consultation is already a requirement under the WHS Act, and it is the only way to see how a system actually lands once it meets a real shift.
4. Keep a competent person owning the decisions the system informs. The software can sort, score and suggest. A person still decides, and still carries the duty.
5. Watch for the SafeWork NSW guidelines. The entry-permit provisions hang off guidelines the regulator has yet to publish, and those will signal how close commencement is.

Where does this sit in the national picture?

NSW is first, and deliberately so. Safe Work Australia, which maintains the model WHS laws the states share, has flagged AI and automated decision-making as an emerging challenge in its national strategy, but it has not made AI-specific model law (Safe Work Australia, 2024). NSW did not wait for harmonisation. The Act even builds in a review, a new section 276D, that triggers if the model laws later cover the same ground.

It is also genuinely contested, which is worth saying plainly. The Australian Industry Group, an employer body, called the Bill "fundamentally flawed and unwarranted", objecting in particular to the union powers to inspect AI systems and to NSW diverging from the rest of the country (Australian Industry Group, February 2026). Unions backed it. Reasonable people disagree on the mechanism. What is not in dispute is the direction: regulators now expect AI at work to be governed like any other source of risk, and the more capable the system, the sharper the question of who owns its risks. That was already the thread in my read of Claude Fable 5 for safety work.

The bottom line

AI belongs around the safety decision, not on it, and NSW has now written a version of that principle into statute. The Digital Work Systems Act does not ban AI, automation or algorithms. It asks a PCBU to manage their risks the way it manages any other hazard, and it makes clear the duty stays with the person, not the platform. Use these systems to allocate, to draft, to analyse, and keep a competent person owning the call, the same test I apply to incident investigation under ICAM and everywhere else AI touches safety work.

The duty is not switched on yet. The work of being ready for it is. If you are mapping how the new law touches your own systems, reach out, I'm always happy to compare notes.