How to Use ChatGPT for Workplace Health and Safety (and What to Avoid)

Most safety managers I talk to already use ChatGPT. They use it quietly, for the writing jobs nobody enjoys, and they are not sure which parts are fine and which parts will get them in trouble. So here is a straight answer. ChatGPT, Claude and Copilot are good at drafting, rewriting and summarising. They are bad at knowing your site, and they are a privacy risk the moment you paste in a worker's name and injury. Everything else in this guide follows from that.

What ChatGPT is good for in WHS work

Use it for the jobs where a rough first draft saves you an hour and a competent person will check the result anyway.

• First drafts of SWMS and JSAs. Ask for the task steps, typical hazards and controls for a job like "replacing roof sheeting on a two-storey commercial building". You get a structure to react to instead of a blank page. Then the crew and a competent person fix it to match your site. Our SWMS guide covers what has to be in one.
• Toolbox talk outlines. "Give me a ten-minute toolbox talk on manual handling for a warehouse crew, with three questions to ask them." Good output, every time.
• Plain-English rewrites. Paste in a procedure that reads like it was written by a lawyer and ask for a version a new starter can follow. This is where these tools shine.
• Summarising long documents. A 90-page code of practice, a supplier's safety data sheet, a regulator's guidance note. Ask for the five things that apply to your work. Then read those five sections yourself.
• Checklists and templates. Inspection checklists, induction outlines, emergency plan sections. Drafts, not finals.
• Thinking partner. "What am I missing in this risk assessment for confined space entry?" It will list things. Some will be wrong. Some will be things you missed.

What to avoid

Do not paste worker health or incident details into a public chatbot

This is the one that matters. An incident report has a name, an injury, often a medical detail and sometimes a witness statement. That is health information. In Australia it is sensitive information under the Privacy Act 1988. In New Zealand it is covered by the Privacy Act 2020 and the Health Information Privacy Code. When you paste it into a consumer chatbot you have disclosed it outside your systems, you do not control where it is stored, and depending on the account settings it may be used for training. If you need an assistant to work with incident data, do it through a system that connects under your own permissions. More on that below.

Do not treat it as a source of law

ChatGPT will quote a WHS Regulation clause number with total confidence and get it wrong. It mixes up Australian states, mixes up Australia and New Zealand, and cites codes of practice that have been replaced. Use it to find where to look, then read the actual regulation or code. For anything that will go in front of a regulator, check the primary source.

Do not let it decide what is notifiable

Whether an incident is notifiable under the WHS Act or the Health and Safety at Work Act 2015 is a decision for a competent person who knows the facts. The assistant can help you read the criteria. It should not make the call.

Do not put an unreviewed draft on site

A SWMS that nobody checked is worse than no SWMS, because it looks like control without being control. Every draft gets reviewed by the people doing the work and signed off by someone competent.

Do not assume it knows your site

It does not know your plant, your traffic management plan, your exclusion zones or the fact that bay three has a blind corner. Everything it writes is generic until you make it specific.

Prompts that work

A few patterns that get better output. Give it a role, a context, a format and a limit.

• "You are a WHS advisor in Queensland. Draft a JSA for changing a tyre on a 20-tonne excavator in a workshop. Table with task step, hazard, risk rating, control, responsible person. Eight to twelve steps."
• "Rewrite this procedure for a reader with English as a second language. Short sentences. Keep every control. Do not add anything."
• "Here is a code of practice. List every requirement that applies to a crew doing hot work in a confined space, with the section number for each."
• "What questions should I ask the crew before we start the first lift with a new crane on this site?"

The safe way to use AI with your actual safety data

Everything above is about drafting and reading. The bigger value is in your data: how many near misses, which actions are overdue, whether LTIFR is moving. ChatGPT cannot see any of that on its own. The unsafe way to give it access is to export the incident register and upload it. The safe way is a safety platform that connects to the assistant under your login.

Safety Space does this with an MCP connector for ChatGPT, Claude and Microsoft Copilot. You ask "which corrective actions are overdue by more than 30 days?" in the assistant, it asks Safety Space for those actions under your permissions, and you get the answer with a table. Only the records needed for the question are sent, nothing is used for training, and an administrator can revoke it. The privacy and permissions page has the full detail. That is the difference between using AI with safety data and leaking safety data to AI.

A simple policy for your team

You do not need a twenty-page AI policy. You need four rules that everyone can remember.

1. Drafts only. Anything AI writes gets reviewed by a competent person before it is used.
2. No names, no injuries, no health details in public chatbots. Ever.
3. Check the law at the source. AI points you to it. It is not the source.
4. If you want AI to work with our safety data, use the connected tools we have approved, not an export.

Put it in the induction, mention it at a toolbox talk, and consult your workers on it. Introducing AI into safety work is a change to how work is done, and both the WHS Act and the Health and Safety at Work Act expect consultation on that.

Frequently asked questions

Can I use ChatGPT for workplace health and safety?

Yes, for drafting and reading work. Not as a source of law, not with worker health details, and never without review.

Is it safe to paste incident reports into ChatGPT?

No. Incident reports contain health information about a named person. Use a connected platform under your own permissions, or strip the names and health details first.

Can ChatGPT write a SWMS?

It can draft the structure. The crew and a competent person make it true for your site. For high-risk construction work in Australia, the SWMS is a legal document, so that review is required.

Can ChatGPT analyse my safety data?

Only if it can see it, and the safe way for it to see it is through a connector that respects your permissions. Safety Space provides one for ChatGPT, Claude and Microsoft Copilot.

Is Claude or Copilot any different?

For this guide, no. The same rules apply to all three. Copilot has the advantage that many organisations already have it under a Microsoft 365 agreement with enterprise data terms, which solves part of the privacy problem for drafting work. It still cannot see your safety data without a connector.