The Trouble With Calling It “Intelligent”

A couple of years ago I wrote a short post telling everyone to calm down about AI. It isn’t intelligent, I said. It doesn’t understand anything. It’s a fancy tool, nothing more.

I’d write it differently now, and not because I’ve been won over by the hype. The trouble was that I’d made the same mistake as the people I was arguing with. We were all using words we hadn’t defined.

Take “intelligent”. In everyday speech it’s a vague compliment that covers everything from solving equations to knowing when to leave a party. Vendors use it to mean “please give us money”. Critics use it to mean “a human mind”, and then point out, correctly, that a language model isn’t one. Both sides are shouting past each other using the same word for different things, which is roughly what happens in most strategy workshops.

“Understands” is worse. When someone says AI understands your business, ask what they’d accept as proof that it didn’t. If there’s no possible answer, the word isn’t describing anything; it’s decorating the sales pitch. But when a critic says AI understands nothing, the same question applies. These systems can summarise a contract, spot a contradiction in a report and explain a joke, sometimes better than the person who forwarded it to you. Whatever you want to call that, “nothing” isn’t quite it.

Then there’s “hallucinates”, the industry’s favourite word for when AI makes things up. It’s a wonderfully forgiving term. It suggests a temporary lapse, like a mind briefly wandering, rather than a system producing confident sentences with no built-in commitment to whether they’re true. A group of philosophers argued in 2024 that Harry Frankfurt’s word for that kind of speech fits rather better than “hallucination”. You can guess which word they meant. I suspect the marketing departments prefer the other one.

What these systems actually do is more interesting than either the hype or the backlash. A large language model has learned the patterns of an enormous amount of human writing, and it produces the most plausible continuation of whatever you give it. Sometimes plausible and true are the same thing. Sometimes they aren’t, and the model has no reliable way of telling you which is which. That’s the part that matters for anyone using it at work.

So here’s what I’d suggest instead of getting excited or getting cross.

Describe what the tool does, not what it is. “It drafts a first version of a report from our notes” is useful. “It’s an intelligent assistant that understands our needs” is a brochure.

Treat fluency as a style, not a guarantee. The better AI gets at sounding right, the more carefully you have to check whether it is right, which is a strange reversal of how we usually judge writing.

And be suspicious of anyone, enthusiast or sceptic, who talks about AI with total certainty. The honest position at the moment is that it’s remarkably capable at some things, unreliable at others, and changing fast enough that today’s confident claim is next year’s embarrassing old blog post.

I should know. I’ve just rewritten mine.