Your AI-Generated Safety Program Is Probably a Liability, Not a Program

A document that reads well and a program that holds up under audit are not the same thing.

Why This Is Spreading Now

Cost pressure explains most of it. A safety consultant costs real money, and a chatbot is free or close to it. Add a genuine belief that AI has closed the gap with a credentialed professional, and the DIY safety program writes itself. Literally.

We've reviewed AI-drafted confined space procedures and fall protection plans for clients that read like they came from an experienced safety professional. Polished language. Confident tone. The right section headings in the right order. The problem doesn't show up at the moment of drafting. It shows up during a COR audit, a ComplyWorks or ISNetworld prequalification review, or after an incident, when a lawyer starts asking pointed questions about what was actually followed on site.

Jurisdiction Blindness and Fabricated Citations

AI language models do not reliably track which rules apply to which employer. Ontario's Occupational Health and Safety Act, British Columbia's Workers Compensation Act, and the federal Canada Labour Code Part II carry different obligations, different penalties, and different tests for due diligence. Ask a general-purpose AI tool to write a safety policy without specifying jurisdiction, and it will often blend requirements from all three, or default to whichever framework shows up most in its training data.

The citations are worse. Ask an AI tool for the specific regulation section behind a claim, and there's a real chance it invents one: a section number that doesn't exist, or a reference to a standard that was superseded years ago. It does this confidently, with no flag that the citation might be wrong. We've seen a "confined space entry procedure" cite an outdated version of the CSA Z1006 confined space standard, word for word, as though it were current.

Generic Hazard Content That Doesn't Match Your Site

A confined space entry procedure has to reflect the actual confined space: the tank geometry, the ventilation setup, the atmospheric testing frequency, the rescue plan for that specific location. AI produces the generic version instead, the one that could apply to any tank anywhere, which means it doesn't actually apply to yours.

Documents That Don't Talk to Each Other

This is the mistake that does the most damage, and it's the one companies notice last. AI drafts each policy, procedure, and form in isolation, one prompt at a time. There's no shared terminology, no cross-referencing, no single source of hazard truth connecting them. Ask it for a fall protection policy on Monday and a working-at-heights procedure on Wednesday, and you'll get two documents that describe the same hazard differently, reference different anchor point requirements, and were never checked against each other.

A real health and safety program is a system, not a folder. The policy references the procedures. The procedures reference the training records and the forms. Everything traces back to the same hazard assessment. COR auditors and ISO 45001 assessors are trained to test for exactly that kind of internal consistency, whether one part of the program actually lines up with another. Contractor prequalification platforms like ComplyWorks and ISNetworld score the same thing when a business submits its program for a bid. A stack of disconnected, AI-drafted documents reads like paperwork, not a program, and reviewers notice fast.

A stack of AI-drafted documents reads like paperwork. Auditors can tell the difference from a real program in the first ten minutes.

No Worker or Joint Health and Safety Committee Consultation Trail

Most Canadian jurisdictions require documented worker or Joint Health and Safety Committee input during program development, and regulators treat that consultation as part of due diligence. AI cannot run that consultation. It can write up a plausible-sounding summary of one that never happened, and that's worse than no summary at all. It creates a paper trail that contradicts reality the moment anyone asks a worker what actually happened.

The Due Diligence Trap

Due diligence has nothing to do with how professional a document reads. It's proof of what was actually identified, communicated to workers, and acted on. A provincial OHS regulator investigating a serious incident (Ontario's Ministry of Labour, WorkSafeBC in BC, Alberta OHS, and so on) doesn't ask whether your policy manual is well written. They ask whether the hazard was identified, whether controls were in place, whether workers knew about them, and whether those controls were actually followed.

This isn't theoretical. In Toronto's 2009 Metron Construction swing stage collapse, four workers died after five of the six riders were not properly trained on the fall protection equipment they were supposed to be wearing, and the platform was loaded well beyond its rated capacity. Metron Construction Inc. was convicted criminally and fined $750,000. A company director was fined separately, in part for failing to make sure non-English-speaking workers received safety material in a language they could actually understand. The paperwork existed. What failed was proof that workers understood it and that the controls it described were actually followed on site.

This is where a credentialed professional's judgment matters most: knowing what actually gets tested, and building the program so it can survive that test. A well-formatted AI transcript has the same problem as a well-organized binder: neither one is actually a safety program.

No Verification Loop

AI gives you a policy shell. It doesn't give you an inspection schedule, a corrective action tracking system, or a trigger for reviewing the program when a regulation changes. A program needs a mechanism for staying current and staying enforced. A document, no matter how well written, is not a mechanism.

What a Credentialed Professional Actually Brings

AI is fine as a drafting tool. The mistake is stopping there. A CRSP brings professional judgment about what regulators, auditors, and insurers actually test, built from tailoring a safety program to real operations instead of templating one. That judgment also comes with accountability: a professional designation with a governing body and a code of conduct behind it, not a chatbot with no liability for what it produces.

Common Questions

Can I use AI to help write my safety program?
Yes, as a starting point. AI can draft language quickly, which is genuinely useful. The problem is treating that draft as the finished program instead of having a credentialed professional review it for jurisdiction accuracy, site-specificity, and internal consistency before it's implemented.

Will COR or ISO 45001 accept an AI-generated safety manual?
Auditors don't check who or what drafted a document. They test whether your program works as a system: documents that reference each other, reflect your actual site conditions, and show evidence of real practice. AI-only output usually fails that test because it produces disconnected fragments, not a system.

Who is legally responsible if an AI-drafted policy turns out to be wrong?
The employer, not the AI tool. In Moffatt v. Air Canada (2024), the BC Civil Resolution Tribunal held Air Canada responsible for inaccurate information its own website chatbot gave a customer, rejecting the airline's argument that the chatbot was a separate entity that should answer for its own mistakes. The same principle applies to a safety program: if an AI-drafted procedure turns out to be wrong, that risk sits with the employer in front of a regulator, not with whichever AI tool drafted it.

Is it risky to paste our hazard data or incident history into a general AI tool?
Yes. General-purpose AI tools can retain what gets typed into them, which is why some companies now ban staff from entering sensitive internal information into these tools. Site-specific hazard assessments, incident investigations, and worker information deserve the same caution you'd apply to any other confidential business record.

What to Do Next

If you've already built a safety program with AI, don't throw it out. Get a credentialed professional to stress-test it: the citations, the site-specific detail, the cross-referencing, the consultation trail, before an auditor, a prequalification reviewer, or a lawyer does it for you.