A WorkSafeNB consultant explains why AI-drafted health and safety documents still need a human check before they go into effect
WorkSafeNB is telling employers that handing health and safety documentation to artificial intelligence does not hand off their legal duty to comply with New Brunswick's occupational health and safety (OHS) laws. The provincial regulator, a Crown corporation responsible for enforcing the province's Occupational Health and Safety Act, published an advisory on September 10, 2026, after its consultants and inspection officers noticed a pattern of employers relying on AI to draft policies, procedures and codes of practice without the expertise to catch what was wrong with them.
John Debly, a health and safety consultant on WorkSafeNB's prevention team, said the pattern surfaced through the team's day-to-day work helping employers build health and safety management systems.
"We noticed the pattern where, anecdotally, from me and my colleagues, employers were placing a lot of faith in AI to craft policies, procedures and codes of practice," Debly said. "In several cases, the employer was doing it with good intent, but they didn't know enough to evaluate what they were getting."
A pattern regulators say employers could not catch
The errors Debly's team encountered ranged from small to potentially dangerous. Some documents cited the wrong section of the Occupational Health and Safety Act and mixed up regulations from one province to another.
"We've seen things like incorrect sections cited, mentioning a section when it should have been a different section of the Act," Debly said. "Wrong information, old standards, confusing American and Canadian standards."
Debly said he has personally had AI tools invent sources outright, including journal articles that did not exist and a fabricated book, complete with a fake ISBN, when he pressed the tool on where its information came from. The risk, he said, is that AI-generated text often reads as authoritative even when it is wrong, which is part of what has made it so easy for employers to skip the verification step. This adds a live compliance example to a broader debate over whether AI should author safety documents at all, suggesting that debate has moved from theory into active enforcement files.
Learning to keep AI on a tight leash
WorkSafeNB is not telling employers to stop using AI. Debly's advice centres on how the tool is prompted and how its output is checked afterward.
"If you're going to use AI, learn how to create a really good prompt," Debly said. "When you ask the question, ask for the references, ask for the section numbers, ask for the links to the references. If you do a proper prompt, you can really keep AI on a tight leash."
Debly recommends anchoring prompts to the New Brunswick Occupational Health and Safety Act and regulations directly, or to the Canadian Centre for Occupational Health and Safety for more general questions, and then having someone with the expertise to evaluate the draft before it becomes policy. Employers without that expertise in-house, he said, can call WorkSafeNB's prevention team before a document goes into effect rather than after an inspection flags a problem. The approach lines up with the guardrails Canadian workplaces need before relying on AI outputs for anything safety-critical.
AI cannot be a legal defence
New Brunswick's occupational health and safety system, like the rest of Canada's, runs on the internal responsibility system: the principle that every party in a workplace, including the employer, shares responsibility for identifying and controlling hazards. Demonstrating due diligence under that system means showing WorkSafeNB that the employer did what a reasonable, informed person would do, and Debly said an AI tool's error does not change that standard.
"Using AI would not in any way ever be able to be used as a legal defence," Debly said. "You can't transfer that responsibility to exercise due diligence to AI, and certainly doing everything a reasonable person would do would be to critically look at what you're getting from AI, or wherever else you get information."
He pointed to safety-critical categories such as fall protection, trenching and sloping angles, hot work and Workplace Hazardous Materials Information System (WHMIS) respirator requirements as examples where an AI-generated error carries the most severe consequences. If an incident investigation later traces a gap back to inaccurate AI-drafted guidance, Debly said, that inaccuracy alone would not satisfy what regulators expect employers to prove during a due diligence review.
Debly's closing message for employers was direct: AI can help draft a procedure, but only a competent employer can make it safe. "AI should support OHS judgment," he said, "not replace proper workplace assessment, worker consultation or employer accountability."
