Monday 4 September 2023

Wittgenstein on AI

Struck by what appears to be a very high degree of conceptual confusion about AI, I've been drawn back to the basic premise of Wittgenstein that the problems of philosophy (or here, "making sense of AI") stem from lack of clarity in the way language is used. Wittgenstein's thoughts on aesthetics come closest to articulating something that might be adapted to the way people react to AI:

"When we make an aesthetic judgement about a thing, we do not just gape at it and say: "Oh! How marvellous!" We distinguish between a person who knows what he is talking about and a person who doesn't. If a person is to admire English poetry, he must know English. Suppose that a Russian who doesn't know English is overwhelmed by a sonnet admitted to be good. We would say that he does not know what is in it. In music this is more pronounced. Suppose there is a person who admires and enjoys what is admitted to be good but can't remember the simplest tunes, doesn't know when the bass comes in, etc. We say he hasn't seen what's in it. We use the phrase 'A man is musical' not so as to call a man musical if he says "Ah!" when a piece of music is played, any more than we call a dog musical if it wags its tail when music is played."

Wittgenstein says that expressions of aesthetic appreciation have their origins as interjections in response to aesthetic phenomena.  The same is true of our judgements to writing produced by AI: we said (perhaps when we first saw it) "Wow!" or "that's amazing". Even after more experience with it, we can laugh at an AI-generated poem or say "Ah!" to a picture. But these interjections are not indicators of understanding. They are more like expressions of surprise at what appears to be "understanding" by a machine. 

In reality, such interjections are a response to what might be described as "noise that appears to make sense". But there is a difference between the judgement of someone who might interject after an AI has returned a result who has a deeper understanding of what is going behind the scenes, and someone who doesn't. One of the problems of our efforts to establish conceptual clarity is that it is very difficult to distinguish the signal "Wow!" from its provenance in the understanding or lack of it in the person making the signal. 

Aesthetic judgement is not simply about saying "lovely" to a particular piece of art. It is about understanding the repertoire of interjections that are possible in response to a vast range of different stimuli. Moreover, it is about having an understanding of the constraints of reaction alongside an understanding of the mechanisms for production of the stimuli in the first place. It is about appreciating  a performance of Beethoven when we also have some appreciation of what it is like to try to play Beethoven. 

Finally, whatever repertoire one has to make judgements, you can find others in the social world with whom you can communicate the structure of your repertoire of reactions to AI. This is about sharing the selection mechanism for your utterances and in so doing articulating a deeper comprehension of the technology between you. 

I'm doing some work at the moment on the dimensionality of these different positions. It seems that this may hold the key for a more rational understanding of the technology and help us to carve a coherent path towards adapting our institutions to it. But in appreciating the dimensionality of these positions, the problem is that the interconnections between the different dimensions breaks. 

It is easy to fake expertise in AI because few understand it deeply. That means it is possible to learn a repertoire of communications about AI without the utterances being grounded in the actual "noise" of the real technology. 

It is also easy to construct new kinds of language game about AI which are divorced from practice, but manage to co-opt existing discourses so as to give those existing discourses some veneer of "relevance". "AI ethics" is probably the worst offender here, but there's lots of words spent of discussing the sociology of "meaning" in AI. 

Equally it is possible to be deeply grounded in the noise of the technology but to find that the concepts arising from this engagement find no resonance with people who have no contact with the technics, or indeed, are in some cases almost impossible to express as signals. 

It is in understanding the dynamics of these problems which is where the dimensionality can help. It is also where experiments to probe the relationship between human communications about the technology and the technology itself can be situated.