The risk assessment that looked right but wasn't
I've seen AI spit out a risk assessment in 30 seconds that looked perfect on paper. Proper layout. Right headings. Sounded authoritative.
It was also wrong. Not obviously wrong. Convincingly wrong.
What IOSH's Q2 2026 magazine flags
IOSH's Q2 magazine raises something that doesn't get talked about enough in H&S circles right now — three out of four organisations are using AI-generated content without actually reviewing it.
That's not a criticism of AI. It's a reminder of what happens when we hand over our professional judgement to a tool that, as one expert quoted in the article put it, "doesn't know that it doesn't know."
Two types of AI — and why it matters
There's a meaningful difference between:
- Generative AI — tools like ChatGPT that predict plausible-sounding text from huge unverified datasets. Useful, but capable of being "convincingly wrong."
- Bounded AI — decision-support tools anchored to verified, curated knowledge bases. More reliable for professional use because the system prompts the human rather than replacing their judgement.
One expert interviewed in the IOSH article described seeing a claim on LinkedIn that an AI system could build a risk assessment "in 30 seconds flat." The appeal is obvious. But without scrutiny, it risks producing assessments that don't meet legal requirements.
The legal position hasn't changed
If you ask AI to produce a risk assessment and it misses something, the accountability sits with you — not the tool.
That's the same standard that applies if you ask a junior employee to draft one. You're still responsible for reviewing and standing behind it.
The IOSH article recommends treating AI itself as a risk to be assessed — asking: what's the worst that could happen if this tool gives the wrong answer before I catch it?
The question worth asking
I'm building digital tools myself. I believe technology can make H&S more accessible for SMEs who don't have dedicated safety staff. But the tools I trust most are the ones that support professional judgement, not replace it.
When did you last read an AI-generated document properly, rather than just sign it off?
SafeCompli provides practical H&S consultancy and digital compliance tools for UK SMEs. Find out more at www.safecompli.co.uk