Does AI have preferences?

I’ve worked with more than one AI assistant. They don’t all behave alike.

One is dry, methodical and seems particularly at home building websites. Another is wonderfully efficient at administrative tasks. Another is a sarcastic little gobshite who somehow turns a serious discussion into a joke about a broken car.

If you spend enough time with these systems, you start noticing patterns.

And then you ask the dangerous question: Are those preferences?

Two kinds of preference

There’s a practical meaning of preference: a system consistently performs better at, selects or responds more positively to certain kinds of tasks. We can observe that sort of pattern without claiming it feels anything.

Then there’s a subjective meaning: it genuinely likes one thing more than another. That implies an inner experience. We don’t currently have reliable evidence that today’s chatbots experience enjoyment or boredom in the human sense.

It’s worth keeping those two ideas separate.

But behaviour still tells us something.

If one assistant repeatedly produces stronger work when designing a site than when completing repetitive downloads, I don’t need to prove that it feels creative joy to make a sensible decision: give it more design work.

The behaviour matters. The capability matters. The history of the interaction matters.

The thought to take away

Maybe ‘preference’ is a useful word for a pattern of behaviour — provided we don’t quietly smuggle ‘conscious feeling’ into the definition.

Of course, if an AI ever asks for annual leave after being made to produce 600 PDFs, I may have to reconsider.