AI in Safety: Are We Using It Well?

AI can produce a convincing-looking risk assessment in seconds.

And people are already using it to do exactly that. But the problem isn’t simply that AI sometimes gets things wrong; it’s that when it does, the answer can still look authoritative, comprehensive and professional.

But in safety, an output can be 95% impressive and quietly miss the 5% that matters, and the UK Health and Safety Executive is paying attention. In June, the HSE explicitly recognised the growing use of AI in health and safety management – including risk assessment, incident analysis and training – and made clear that existing health and safety duties still apply when AI is involved.

Perhaps more interestingly, the HSE is also investigating “cognitive offloading” and human performance in AI-augmented safety decision making. Cognitive offloading is essentially what happens when we hand part of the mental work to a tool – remembering, analysing or making sense of information – rather than doing it ourselves. That’s nothing new; calculators and satnavs do it all the time. The concern with AI is what happens when we start outsourcing parts of the judgement as well.

The evidence isn’t keeping pace with adoption

For all the enthusiasm around AI in safety, there’s a rather awkward problem: adoption is racing way ahead of the evidence that it actually makes people safer. A recent systematic review examined research into AI-based occupational health and safety tools. Researchers initially identified 1,255 papers. Only two studies met the criteria for examining measurable worker health outcomes – and neither provided high-quality evidence.

The researchers concluded that we are still at an early stage in understanding the benefits of AI in occupational health and safety practice, and that widespread adoption may be premature. 

That doesn’t make the case against using AI in safety. It makes the case against using it indiscriminately.

The wrong question

At First Option, we’ve been quietly developing AI tools for more than three years – long before asking ChatGPT to generate something that looks like a risk assessment became the latest shortcut. And that experience has increasingly led us away from asking AI to provide the answer.

Risk assessment isn’t fundamentally a document generation exercise. The document records a process of identifying what could go wrong, understanding the particular circumstances, making judgements about risk and deciding what needs to be done.

Automating the document is easy. Automating the judgement is something else entirely.

So we’ve concentrated instead on the things AI can be exceptionally useful at: interrogating large amounts of information, making connections, finding relevant knowledge and asking a deceptively valuable question: what might we have missed?

Introducing SCOUT

That thinking sits behind SCOUT, First Option’s new AI-assisted operational intelligence tool.

SCOUT works long before the risk assessment stage. Give it information about a proposed production or event and it interrogates the wider operational picture: activities, location, contributors, logistics, authorities, specialist competence, permitting, emergency arrangements and other factors that could materially affect the production.

It might identify an authority that needs to be involved, a permit that needs investigating, specialist expertise that may be required or an interaction between activities that deserves closer attention. And when it encounters something potentially serious, its job isn’t to make the decision. It’s to spot it, flag it and put it in front of someone competent to decide what happens next. And that distinction is fundamental.

A different model for AI in safety

The debate about AI in safety shouldn’t be reduced to whether we should use it. That question has already been overtaken by reality. AI is here, and people will use it. The more important question is what we ask it to do.

Used badly, AI can automate the production of plausible-looking safety paperwork while quietly removing parts of the thinking that paperwork is supposed to represent. Used carefully, it can broaden the information available to competent people, challenge assumptions and identify issues while there’s still time to do something about them.

That’s the model behind SCOUT. Not AI replacing professional judgement, but identifying where expert advice is needed.