There is a steady stream of advice at the moment telling us to stop saying that an AI knows, wants, thinks, decides or tries. One recent example sets out how to talk about AI without adding to the hype. The worry behind it is fair enough: a reader who takes them at face value ends up believing something false about what the software is.

But the cure is harder to swallow than it looks, because the words are not a decoration we added to English and can now peel off. Wittgenstein suggests that a word means what it does in the games we play with it, and there is no core definition sitting underneath waiting to be enforced. We already say that a car is struggling, that the washing machine has finished, that the boiler has given up. Nobody is confused. All that is happening now is that usages like "aware of", "trying to" and "perceiving" are splitting and expanding in new and interesting ways.

So here is what the advice sounds like when someone actually follows it, on a machine nobody has ever accused of being intelligent.

A: The car won't start. B: Mm. Won't, though. A: Sorry? B: Won't means it could have gone the other way and chose not to. Sorry. I'm trying to de-anthropomorphise my speech. I read a piece about how we talk about AI and now I can't turn it off. A: The car doesn't start, then. It hates the cold. B: Ye-es. Petrol evaporates less readily when it's cold, battery has less to give, nobody in the situation hates anything. Harder than it looks, this. A: It turns over and then gives up. B: Giving up is what you do after trying, and it hasn't been trying, it's been rotating. Aargh. I know how I sound. A: Listen to it. That's a car in distress. B: Solenoid clicking, six or seven times a second, which is what they do when the voltage sags. Although I grant you it sounds precisely like distress, which is the whole difficulty and I have no answer to it. A: Can you look at it or not. B: I have. Eleven point one volts. It wants... no. Gosh. It is below the voltage the starter needs. A: Is it dead? B: No. Never was. That one I'm holding the line on. A: So what killed it? B: You left the interior light on since Thursday. A: So it's my fault, not the car's. B: Yes. That's the thing I've been failing to say for ten minutes. A: It's a good old car. Been good to me. B: It's been reliable. A: Same thing. B: It is not the same thing. (pause) It is a bit the same thing. Aargh.

B is right about the physics every single time and is no use to anybody. A gets the car going.

None of this licenses the context-free claim that a car has feelings or that a large language model is a person. It is the narrower point I make in Yes, it's ok to say that an AI understands what you say and in Scare quotes, the Turing test, and memory: the in-context uses are unavoidable, they are not confused, and policing them buys less clarity than it costs. Save the argument for the claims that actually mislead.

Posted on LinkedIn. Dialogue written by Claude Code, which is either the perfect author for it or the worst one.