The Capable Junior

A useful way to think about AI is as a very capable junior assistant.

It can do an enormous amount of work very quickly. But imagine giving a junior member of your team half the information about a project, some fairly vague instructions and no opportunity to ask questions – then treating whatever they produce as the finished answer. You wouldn’t blame the junior for not knowing something you never told them.

And you certainly wouldn’t assume that a beautifully written document meant they’d somehow discovered all the missing information for themselves. AI is much the same.

The quality of what you get depends enormously on the information, instructions and context you give it. The difference is that AI can be extraordinarily good at disguising those gaps. Rather than coming back and saying, “I don’t really understand the job – can we talk about it?”, it will often make the best of what you’ve given it and produce something that sounds remarkably convincing. That’s why the instructions you give it matter so much.

Give it your rough notes from a recce and ask it to organise them into something coherent. Ask it to summarise a long piece of guidance. Give it a plan and ask what questions you might need to consider. Use it as a second pair of eyes, a sounding board or a devil’s advocate. One of the most useful questions you can ask an AI is simply: what have I missed?

Those are all examples of AI supporting the thinking rather than replacing it.

And there’s an important distinction: asking an AI “what might I need to think about?” leaves you in control of the decision. Asking it “is this safe?” hands it a judgement it isn’t equipped to make. The same applies to risk assessment. Using AI to suggest possible hazards for you to consider can be genuinely useful. Asking it to decide which hazards are significant, what controls are adequate and whether the resulting activity is acceptable is something quite different.

There are other warning signs. Be cautious when AI produces risk scores without enough information to justify them. Be cautious when it gives definitive answers about legislation, compliance or specialist requirements without citing reliable sources. And be “particularly” cautious when it confidently fills gaps in information, rather than recognising that those gaps need to be resolved.

None of this means AI has no place in safety. Quite the opposite.

Used well, it can help organise information, accelerate research, challenge assumptions, identify possible omissions and turn messy information into something much easier to work with. The trick is to use AI for the things it’s genuinely good at, while keeping the human and professional judgement where judgement matters.

A simple rule of thumb is: Use AI to support your thinking. Don’t outsource the judgement.

There is one other precaution worth mentioning. Safety work can involve information you shouldn’t casually put into an AI system: personal or medical information, incident details, security arrangements, confidential project information or details about contributors and crew.

Different AI systems have different arrangements for how information is stored, processed and used. So before putting anything sensitive into an AI tool, make sure you understand what system you’re using, what it might do with whatever you give it, and whether you’re authorised to put that information in there in the first place.

Or, more simply: Protect the input. Check the output. Own the judgement.

But using AI well isn’t just about the person sitting in front of it. It also depends on how the AI system itself has been designed – what information it works from, what it’s allowed to do, and what else it might be doing with what you feed into it.

And that’s where AI in safety starts to get much more interesting.