Can AI make aesthetic judgments?

Question

I'm guessing AI could maybe be used to determine the simplest solution, and I think it can clearly be used to check when a proof works. But can AI make aesthetic judgments? If so, then it may well end up revolutionising art, right? If not, all of its creative work will be missing something: self criticism too.

Answer 1

It depends on what you mean by aesthetic judgements. You can certainly program an AI to give verdicts based on parameters, and the parameters can be related to the characteristics of music, images, objects etc that we would take into account when making aesthetic judgements, so in that sense, yes, AI can take judgements about the aesthetic merits of particular subjects. However, if you mean can an AI make judgements which are grounded in its own emotional response to a subject, then the answer is no.

Answer 2

I'm guessing AI could maybe be used to determine the simplest solution, and I think it can clearly be used to check when a proof works. But can AI make aesthetic judgments? If so, then it may well end up revolutionising art, right? If not, all of its creative work will be missing something: self criticism too.

Good question. However, first let's demystify what AI does. A classical computer is much better at checking when a proof works assuming that it's given the rules. What's referred to as AI nowadays, namely deep learning, does this in a completely different way. It does it by learning the rule all by itself. However we have no way of checking whether it actually knows the rule except to judge it by its output. This is quantified through some error function, and the optimization algorithm (the way AI learns) will try to minimize the error to its lowest possible value.

However, depending on the network architecture, deep nets can be good or bad at different tasks. Until very recently, architectures have been tailored to the type of data input e.g. natural language, images, sound, video etc. This is still mostly the case, except that a relatively general architecture has emerged called a Transformer, that's more or less generalizable across architectures.

When it comes to aesthetic judgement, AI learns pretty much the same it way it learns proofs. That is, it tries to generalize the underlying rule given some pattern. And to make deep nets work they have to be fed very, very large training samples. So what will happen is that it will be fed a catalogue of the entire history of human art, let's say, and then it will be asked to produce an artwork.

Now consider how a human's aesthetic judgement is developed. It's developed through exposure to prior art, where their individual preferences evolve against some cultural milieu. In other words, the apple doesn't fall too far from the tree. We didn't get Picasso after Leonardo Da Vinci, we got him after the post-impressionists and Fauvists, and he himself went through those phases. And against the milieu of a new physics, he moved beyond mimesis and experimented with simultaneity of spatial dimensions. From the Renaissance we got the development of perspective, right as mathematics was breaking new ground and the scientific revolution was dawning.

A deep net is in principle emulating this process. It's given some prior information, and it's building from there. Human originality works in much the same way. But given how big an AI's prior is in comparison to a human, it may even generate artistic outputs that far exceed the originality of a single human.

However, while the broad strokes are similar, there are a whole host of differences wherein current AI and humans fundamentally differ. One is that the human can do ingenious things and even shift the frame of their activity with a very small prior, whereas AI needs gigantic datasets to be fed to it. Partly this is because a living organism integrates information far more deeply than an AI does and we don't truly understand the higher levels of that integrative process in the brain (we know from the vision system that it's partly hierarchical i.e. assemblies of specialized neurons will detect layers of features which are integrated into dynamic images, which is what computer vision is based on via convolutional nets).

Finally, an AI does not yet have a virtual self, which is a form of informational integration that gives the organism a view of itself from the inside. It does not have intentions the way a human does. It doesn't know that it knows. It cannot reflect back on itself. AI researchers don't like these philosophical points, as they think they're woo woo, but I think they're deeply mistaken.

So to answer your question, the difference between checking proofs and aesthetic choice is not as wide as it prima facie appears in current AI. Once you understand the network architecture behind AI, you can see that it tries to emulate the learning capabilities of humans, but its "representations" are way, way too coarse to have anything like intentionality, "intelligence", justify its claims to knowledge, and even have experience affect/sensations, which constitutes the wellspring of art. I didn't talk as much about the latter -- pattern recognition is fine, but human generated art is motivated by affect, something that AI doesn't have and probably will not have for a long time. The closest to affect is what's called reinforcement learning, which is a way of nudging a network through negative and positive reinforcement toward some desired output. Human affect works partly like this in practice, but without the actual what's it's likeness or qualia, mathematical reinforcement doesn't come close to emulating the internal states of living organisms.

Answer 3

In theory, AI might be able to identify a Picasso, same as it might imitate one, better than any human expert, though I gather that at the moment it is poor at identifying itself. It then might well be able to write some art criticism on what it does. And I'm sure some people already prefer midjourney to Picasso, and part of why they are arguably wrong might involve analogy with how photography is not "better" than a painting.

So I would think that AI is always going to be doing something different to human artists, not just how it works (as in competing at Go or AI hacking), but actively playing a different game. There's no clear cut answer as to whether an army of LLM "artists" can outdo all of humanity, without defining the value of art, but I would consider it impossible according my own nebulous sense of that, and the same might even be said of philosophy.

Answer 4

We are incapable of explaining why art is valuable to a philistine, and all AI starts out that way, so while it may be able to emulate people's tastes (spotify does a job on this already), it will be unable to do so critically, because even we don't know why is valuable.

I suppose it could be fed certain hypothetical scenarios (such as, because art is free creative activity), but we don't know that's why (it is an approximation to a pre-theoretic 'faith' or similar), so the machine won't "know" either.

Answer 5

(Note that this is about AI of today: they only do number crunching and computational logic.)

I say no, AI cannot make aesthetic judgments. It can imitate it, though.

There is no universal aesthetic judgment, so the AI must be trained according to someone's taste. It can learn to associate numerically expressible features to someone's likes and dislikes, but even then the AI is not making aesthetic judgments per se.

The same AI could be trained in an entirely different manner to do entirely different classifications. Eg. whether the painting has bunnies or not. The processing would be fundamentally the same in both cases. The only difference of any significance is the meaning we assign to the output: "this means that the painting is probably pleasant" is not fundamentally different from "this means that there are probably no bunnies in the painting".

Because subjectively (to a human being) aesthetic judgments are fundamentally different from spotting whether there are bunnies in the picture or not, and to an AI they are the same, I argue that the AI cannot do bona fide aesthetic judgments. It only does clever computations that can produce surprisingly accurate classifications, but the underlying process does not involve anything that is actually about aesthetics.

In short: the AI does not know if something is pleasant. It only can tell if there are features that correlate with someone's taste.