The Risk Assessment Trap

So let’s put this into practice. Try this for yourself.

Imagine you’re filming a simple interview beside a lake. A small crew, a contributor, two cameras and some lighting. You ask an AI to identify the significant risks. And it will probably do a pretty good job; it might identify slips and trips, working near water, electrical equipment, manual handling, weather, cables and perhaps members of the public. All perfectly sensible.

Now let’s add one more piece of information. The interview is taking place at night. The risk picture changes. Lighting. Visibility. And access and emergency arrangements may suddenly become much more important. Now add another. The contributor is fifteen years old. Again, the picture changes.

And another. Getting to the location involves carrying the equipment down a steep 400-metre path. And finally, one of the cameras is going to be operated from a small boat.

By now, we’re describing a substantially different scenario from the one the AI originally considered.

But here’s the important point. The AI wasn’t necessarily wrong the first time.

It identified risks from the information it had been given. The problem was the information it hadn’t been given.

And that’s a fundamental weakness in the idea that you can simply describe an activity, press a button and generate a risk assessment. A good risk assessment isn’t just about producing a list of hazards and controls. It’s also about understanding the activity well enough to know what questions need asking.

What’s actually happening? Where? Who’s involved? What equipment is being used? What changes if something goes wrong? Is there anything unusual about the location, the people, the activity or the way it’s being done?

Sometimes the most important information isn’t in the original description at all.

A well-designed AI system can help you explore those questions. It can prompt you for more information, identify possible gaps and help you think about things you may have overlooked. But simply generating a more comprehensive-looking document “doesn’t” solve the underlying problem; in fact, it can make it harder to see.

Because every missing piece of information at the beginning can flow through into the hazards, the controls and ultimately the finished assessment, while the document itself still looks perfectly convincing.

So the problem with “generate me a risk assessment” isn’t really the AI. It’s the idea that generating the document is the same thing as doing the assessment. It isn’t.

But that doesn’t mean AI has no place in the process. Quite the opposite. The trick is knowing which parts of the work it can genuinely make better – and which parts we shouldn’t hand over to it at all.