Intentional stance

You are playing chess against a computer. To guess its next move you do not trace the electrons, and you do not read the program. You ask what it is trying to do, what it knows about the position, and what a sensible player would do with those aims and that knowledge. Daniel Dennett called this taking the intentional stance towards the machine.

Three stances

Dennett distinguished three ways of predicting what something will do.

• From the physical stance you use the laws of physics and the thing’s physical state. It always applies and is usually impractical.
• From the design stance you assume the thing will work as designed. You predict that the alarm will ring at seven without knowing its circuitry.
• From the intentional stance you treat the thing as an agent, attribute to it the beliefs and desires it ought to have given its situation, and predict that it will act rationally on them.

Each step up gives away detail and buys economy. The intentional stance is risky, since the thing may not be rational, and for complex systems it is often the only method that works.

What it claims about minds

Dennett’s thesis is that there is nothing more to having beliefs than being a system whose behaviour is reliably predicted in this way. To be a believer is to be an “intentional system”. That is not, he held, the view that beliefs are useful fictions: the patterns the stance picks out are really there in the behaviour, and someone who ignored them would miss something true.

Critics think this leaves out what matters. On their view a system has beliefs in virtue of its inner organisation, and behaviour that merely invites the description is not enough.

Language models

The stance has become a practical question. Chatbots are far easier to predict as agents with beliefs and aims than in any other way, and that is how people use them. Murray Shanahan urges care: a language model is at bottom a system that continues text, and saying that it “knows” or “thinks” something is a convenience that readily turns into anthropomorphism.

Where I stand

I think Dennett is right that the stance earns its keep by predicting, and that its success with chatbots is real. But predicting a system is one purpose and learning from it is another. When I ask whether a model’s answer can give me knowledge, what matters is how the answer was produced: whether its truth played any part in why it was given. The intentional stance is silent on that. A system can be well predicted as “believing” what it says while the process behind the saying is indifferent to whether it is so.