Can AI write a Safe Work Method Statement (SWMS)?

It can draft one quickly from a good prompt, which saves time on structure and wording. But a SWMS has to reflect the actual task, plant, people and site conditions, and a competent person must review and sign off. Use AI to draft, never to decide.

Does WHS legislation allow the use of AI tools?

Nothing stops you using AI as an aid. What you cannot do is transfer your duty to it. Under the model WHS Act the primary duty of care rests with the PCBU and duties are not transferable, so the accountable person stays accountable no matter which tool produced the output.

What are the safest first uses of AI in safety?

The lowest-risk starts are drafting (toolbox talks, procedure rewrites), searching your own documentation, and first-pass analysis of incident and leading-indicator data. In each, a competent person still reviews the output, so a wrong draft costs minutes rather than exposure.

How do I stop AI hallucinating about safety law?

Don't rely on the model's memory. Ground it in the primary source, the actual Act, regulation, code of practice or standard, and verify any clause or section it cites against that source. A model reasoning from the text is far safer than one recalling it.

Will AI replace safety professionals?

No. It changes the work. The routine drafting, searching and first-pass analysis compress, which frees time for the judgement, relationships and field presence that actually move safety outcomes, and that AI cannot do.

AI in Workplace Safety: A Practitioner's Field Guide (WHS & OHS)

Artificial intelligence is genuinely useful in workplace health and safety: for finding patterns in incident and leading-indicator data, for drafting and summarising, and for making procedures searchable. It is genuinely dangerous the moment it is trusted to reason about causation, to quote regulation it has not read, or to make a call a competent person should own. The whole skill is telling those two apart.

I have spent a decade across WHS, quality and environmental work. I have investigated critical incidents as an ICAM lead investigator, run Zero Harm performance and programs for a workforce of around 50,000, and lately built the predictive analytics and applied AI that sit underneath it. So this is not a vendor's pitch. It is the field guide I wish I had when I first pointed these tools at safety problems.

Where does AI actually help?

Start with the uses where a human stays on the decision and AI does the heavy lifting around it. Three hold up under real conditions.

Finding patterns across large datasets

AI reads more rows than a person ever will. Pointed at incident records, hazard reports and leading-indicator data, it surfaces clusters, recurring contributing factors, and the chronically under-verified risks that hide in the averages. On a Power BI semantic model spanning tens of thousands of workers, I have built rolling 12-month LTIFR and TRIFR forecasting and a critical-risk lookahead that flags exactly those blind spots. The model will not tell you why something is happening. It tells you where to look, and that is where most safety analysis stalls.

Drafting and summarising

Toolbox talks, investigation narratives, procedure rewrites, board summaries: the first draft is where the hours disappear. AI collapses that. A strong prompt produces a structured starting point in seconds, and your time goes into the judgement and accuracy that only you can add.

Triage and search

Most organisations hold more safety documentation than anyone can keep in their head. AI turns that pile into something you can question directly: what does our permit-to-work procedure say about isolation? The answer still needs checking against the source, but the search is no longer the bottleneck.

Where AI gets dangerous

The failures share a shape too: the tool is trusted to do something only a competent, accountable human should. Four show up again and again.

Causation, not correlation. AI finds what co-occurs. It does not understand why an incident happened. Hand it an investigation and it will produce a plausible story, which is exactly the risk, because plausible is not the same as correct.
Confident, wrong regulation. General models will quote a clause, a section number or a standard that does not say what they claim. In a safety context that is not a quirk; it is a hazard. Verify against the primary source, every time.
Automation complacency. The better the tool looks, the less people check it. Designing for a human to stay genuinely in the loop, not just rubber-stamping, is a control in its own right.
The accountability gap. A tool cannot hold a duty. If something goes wrong, "the AI said so" is not a defence and never will be.

A four-question test before you adopt it

Before you put AI into a safety workflow, ask:

What would a confident wrong answer cost here? Low, like a rewritten toolbox talk? Or high, like a risk assessment someone relies on?
Who verifies the output, and are they competent to? If the honest answer is "no one," stop.
Can the output be checked against a primary source? The Act, the regulation, the standard, your own data.
Is a person still making the decision that carries the duty? They have to be.

Clear that test and the use is usually safe to adopt. Fail it and you are not automating work. You are automating risk.

How do you adopt AI without losing rigour?

Keep a competent human on the decisions that matter. Ground every regulatory question in primary legislation rather than the model's memory. Validate outputs against your own controls before they reach the field. And govern the whole thing: decide where AI is allowed, who reviews it, and how you would explain its use to a regulator. Done that way, AI does not dilute professional rigour. It gives you more time to apply it.

The principle that does the most work here is grounding: making the model reason from the actual law instead of a fuzzy recollection of it. For a worked example, see how I encoded the WHS Act 2011 and HSWA 2015 into an AI skill.