Putting AI In Its Place

So far, we’ve mostly talked about AI as something you interact with directly.

You give it information, ask a question and get an answer. But that’s only one way of using the technology. AI can also sit inside a much larger system, where its role is deliberately limited. And in safety, that distinction matters. Think back to our risk assessment example. The obvious approach is to give an AI a description of an activity, and ask it to generate the hazards and controls. It’s quick. It’s impressive. But as we’ve seen, it can also conceal some pretty fundamental weaknesses.

A different approach is to decide which parts of the process AI is actually good at – and constrain it to those tasks only. For example, the core safety guidance might come from a library developed and maintained by competent safety professionals rather than being invented by an AI each time.

AI can then be used to help identify which parts of that professionally-developed knowledge might be relevant to the activity you’ve described. It can look for possible omissions or inconsistencies. It can compare what you’ve said you’re doing with the controls you’ve selected. It can help identify questions that may still need answering. And it can point towards external information or specialist input that may be worth considering. That’s a very “different” role from asking AI to be the safety adviser.

The AI is still doing something valuable. In fact, it’s doing exactly the sort of things computers are extraordinarily good at: processing lots of information quickly, finding relationships, comparing information and drawing attention to things that might otherwise be missed. But in this case the system has been designed around the limitations of the technology, rather than simply hoping those limitations won’t matter.

And importantly, there should still be a human at the end of the process. Someone needs to decide whether the information is “relevant”, whether the controls actually work in the real world, and whether the finished assessment reflects what’s actually going on.

So when you’re evaluating any AI safety tool, don’t just ask does it use AI – that tells you remarkably little. Ask a more useful question is what is the AI actually being allowed to do?

Is it generating safety advice from scratch? Is it working from professionally developed information? Is it helping to find gaps and ask questions? And where does professional judgement enter the process?

What matters isn’t whether a safety system uses AI. It’s the role AI has been given within it.

And there’s another interesting possibility. What if, instead of waiting until you’re ready to write a risk assessment, you could use AI much earlier – while you’re still trying to understand what the project might involve. That’s where we’ll go next.