Re: Can Computers Create Art?

Last updated 2025- 3- 5, 4: 8:13 A.M. UTC+00:00

I'm half crazy, all for the love of you!

Regarding content of the talk itself

The speaker is known to have given the same talk over the years, but he has definitely been updating the set of slides, there was a slide with Miyazaki Hayao in Ghibli-style that you would expect to see on Twitter right now.

The speaker brushes off text-to-image generation as "just a fad" which I feel isn't exactly appropriate, though it doesn't contradict his message. Since this talk is more philosophical in nature, maybe it's fine that he doesn't address the current realities of industrialization or commercialization.

It seems he specifically avoided mentioning some less righteous aspects (such as the industrial reality). But I think the main arguments are generally valid, and the message is encouraging. I will copy them here:

1. Software offers technologies for people to make art.
2. Computers should not be considered artists.
3. Making and understanding art will change.
4. Historical technologies provide many useful lessons.

Who should be considered an artist?

Modern machine learning models really require collecting large swaths of data. The contributions of any specific image in the training data in the resulting inference result is miniscule, yet with the large quantity of data they add up to something substantial. With the Spiderman credits example, the position of the speaker seems to suggest that everyone who contributes in the training dataset is inherently an artist behind any generated work, as the effect of the contribution of any single VFX artist is vanishingly small in the final product. I'm sympathetic to this view, despite how little to no credit they receive in actuality.

In the AARON (Harold Cohen) example, the speaker seems to suggest that those who contribute to making the result possible are also inherently an artist, e.g. to program the system and software.

Indeed, the speaker uses Pixar as an example of a highly integrated environment between the development of the tools and the artists using them. It becomes difficult to say, within the work put into making the product possible, where the development of the tool ends and where the creative activity of making the work begins. In the current day it's even harder, and many would argue programming itself is also a creative activity. For these people, their vision guides the approach to making their tools, I've previously written about this dichotomy here.

In the current age where AI contributions are essentially inescapable, coupled with the Rousseauean view on any artist's living state, it seems like trying to draw clear cut boundary using this rule leads to an infectious model, where everyone eventually becomes an artist of everything that is made. For insatnce, because everyone contributes to the dataset of LLMs, and everyone is being influenced by the output of LLMs, which in turn influence and output of any speaker, despite how vanishingly small that contribution is presented in the output (that output then in turn get gobbled up in the training set, and so on).

Of course, in the real world it's unrealistic to be able to credit everyone and provide appropriate compensation. Suppose 500 years ago Bob set a big fire in the forest, and that lead to a specific barren look of the landscape today which I, as a photographer, find fascinating, I then take a photo with my phone's camera, isn't Bob also inherently involved in my creative process? Indeed, when this happens in the real world, we simply relegate to saying "mother nature is really beautiful". To take this view to the extreme is the concept "Everything is a Remix", in the absence of an arbitrary threshold.

So perhaps, the computer is not an artist, it's those who built the hardware and software and made everything possible who contributed to the creative process of the end users creating art with computers. In any case, it still isn't possible right now for an AI to adapt and evolve on its own without a human mechanism (reinforcement learning).

I'm happy with this view, but it also means what it exactly means to be an artist of a work is becoming less meaningful, and the author addresses this point by deferring to a social view of art, where we place more emphasis on the relationships and stories behind a piece of work. It seems like a necessary development.

The speaker also doesn't address ethics in data collection and compensation (or the lack thereof), which is fine, since it doesn't really change whether a computer can be an artist. Indeed I think not compensating is highly uncool when our connections are closer than ever in the age of the Internet, but it doesn't change the nature of the work like in some 'fruit of the poisonous tree' argument, as in, something suddenly stops being art solely because of it.

We kind of don't care what art is?

As the speaker says, our understanding of art will continue to evolve, as it had since forever. The speaker proceeds to posit something along this line of thinking:

If you have to ask whether something is art, it probably is.

Indeed, it isn't the important question. Quoting a slide from the talk:

The real questions:
Does it create emotional connections?
Does it have meaning? Is it original?
Is it ethical? Is it honest? Is it aesthetically valuable?

Yes, the responses to these questions will change. But I agree in that they offer a good starting point to understand our relationship with the understanding of art in this era. These considerations are almost orthogonal to whether something should be considered art, at least in our current era.

Responses to audience questions

Someone brought up the issue of corporations being categorized as persons in the US, and wonders if it's possible that the same happens to machines. His response was "you don't have to believe what the US supreme court says!", which I think is slightly funny but ultimately a little depressing because of the way affairs are affected by legislation and government action, whose motivations do not always align with our understanding of concepts.

A different view, not from the talk or the speaker himself, about substituting social interactions by monetary transactions, is a little odd in my opinion. I think neither is necessary or sufficient for the other, so I can't really say much to this view.

Aside

If you believe that creative endeavours are NP-hard, perhaps there is some conclusion we can say about where we stand today about AI generated content. Let's consider the activity of writing a story in text-form, for instance, since so many creative endeavours involve telling stories. The architecture of models in deep networks is almost always fixed, and generating text, either autoregressively or not, involves a fixed number of steps (either length or refinment iterations, depending on the type of model). So generation is polynomial time.

Given P vs. NP, where the majority opinion is in the negative, I think it's quite unlikely that ML algorithms be able to consistently generate 'the best' stories (with regards to any reasonably fixed metric), because it would mean we can solve NP hard problems in polynomial time (by the current sort of models that we see, LLMs, for instance). I think this specific view is highly researhced and published already.

With that said, we don't live in a vacuum and neither do the ML models, for most people it's really about 'good enough' and not 'the best'. And indeed for this kind of deliverables, AI is already being used to replace human workers, I think this has always been a forseeable development, though, since learning algorithms are mode-seeking (see mode collapse). The industrialization of the technology is sad in a way, but it still has nothing to do whether a computer is an artist within itself, of course!