Chernobyl’s nuclear accident was a defining event for many people of a certain age. As a boy, I literally lost sleep over my worry that France would suffer a plague of radiation sickness when the toxic clouds blew over us, raining death from the skies.

Happily, the radioactive cloud officially stopped at the French border. That, at least, is what the government told us. In response my father (wisely but unhelpfully, in retrospect) simply instructed me not to believe everything I read.

If it was true then, how about now? Back when almost everything printed was vetted, verified, fact-checked, authorised, and cleared by a legal team before setting the type and running the presses, we could believe a lot more of it.

Today is different. Almost everything we read is unverified, to the point that the ludicrous idea of alternate facts has gained political traction. A great deal of what we read has been published by people with no vetting at all. We’re fed all sorts of old cobblers.

The Internet is an entirely ungoverned space for the publication of ideas ranging from the scientifically proven to the utterly nonsensical. Unfortunately, it is also where Large Language Models – those doyens of the AI landscape – learned everything they know.

Someone early in their career made a LinkedIn post the other day about an irony she’d spotted. She’d always been told in school that Wikipedia is not a reliable source. Yet today, in the office, her manager frequently asks her to find answers with ChatGPT. “How’s that different?” her post mused.

We’ve become incredibly easy to please when it comes to AI-generated information. Writing in the New York Times, Anne-Laure Le Cunff reported that more than 60% of Google searches by Americans now end without a single click. The search engine has become an instant-answers machine. Even reading a Wikipedia article is often too much effort in the AI age, when a machine will do it for you for free.

Le Cunff argued this trend is “undermining curiosity.” By jumping straight to the answer, the shortcut is “threatening our ability to understand the world.” Whilst those dramatic phrases tell only part of the truth, I have no doubt that, rather than making us more knowledgeable, the instant gratification delivered through an AI search window eliminates context and background, and short-circuits our ability to learn.

That can’t be good, right?

As we replace underwriting assistants with bot operators and prompt writers, we should think about the future. One learns how to underwrite by thinking it through, not by jumping straight to the rate. Understanding all the moving parts of a risk and how they interrelate with each other is a critical skill for any underwriter, their foremost essential ability. Dramatically curbing the early-career research and analysis process in favour of solutions delivered by internet-trained Large Language Models could be catastrophic for our industry, especially given the existing talent-pipeline problem.

But no one is using AI in that way, right (hint: if you are, don’t admit it).

You can ask Claude how many barbers in Chicago, then gamble on an answer delivered in less than a second. Alternatively, you can use AI as a tool to spot trends, distil insights, and perform information-intensive tasks that would otherwise take hundreds of hours. That’s what it’s for, and best at.

Consider this. Piracy in the Gulf of Aden, off the Horn of Africa, took a massive upsurge about 20 years ago, when I was a K&R underwriter. My colleagues and I scrambled to come up with a way of rating the seafarers’ capture and ransom risk that was suddenly bearing down on vessel owners. We worked on it tirelessly for an absolute age.

The eureka moment came when we realised that the pirates’ boats had a top speed of about 10-11 knots (maybe 12 MPH), and that their ladders were about three metres long. That meant that any boat that was faster, or had a freeboard greater than about four metres, was pretty much in the clear. If they had both, they were probably untouchable. Those were the good writes, and we hoovered them up. It was a very crude risk selection process, but it worked.

Nowadays we may very well have come up with the same answer (and very much faster) by enlisting AI to help solve the problem. But simply asking it how to rate a K&R risk in piratical waters would no doubt have generated some colourful answers, but not the one we wanted.

The trick is to guide the bot carefully towards the solution. We could begin, for example, by asking for a list of ways to prevent people from boarding boats without authorisation. Physical barriers are very likely to be suggested. We could then ask for a list of physical barriers.

When I asked ChatGPT exactly this series of questions, the first answer to question one was “physical barriers.” The first to the second was “high freeboard.” We would still have to do some active thinking, but from there, much less distance is left to travel.

AI is a fantastic tool to enhance the work of a curious, expert individual. However, you rely on it blindly at your peril. Not only does it sometimes base its instant answers on alternative facts it learned on the internet. Sometimes it just makes things up. If you want more information about that, I refer you to the Wikipedia page Hallucination [artificial intelligence]. You can probably believe it, but if it seems a bit long, just ask Claude for a summary.

* Like every Insurance Technology Diary entry about AI, this one is accurate only to the best of my knowledge at the time of writing. The pace of AI progress is so great that I cannot guarantee it remains so now that it’s finished, let alone when you read it.

* I’m setting down my Diary for a summer break now. I will be back in the run-up to conference season.

Guillaume Bonnissent is CEO of Quotech.