Can AI produce original, compelling work?
It's definitely true that generative models are capable of producing results, that very closely resemble specific examples in the training data. But can we assert that they do so in every instance?
On the other hand, can we assert the opposite: that the models are capable of producing results that don't closely resemble any example in the training data?
We can't make the models generate everything they can and compare it all to the training data. And given any instance of generated material, we can't check the entire training dataset to make sure that no closely resembling examples exist.
All we can do is rely on theory about how the models work.
Models form features of the training data at different levels of abstraction.
For example, it probably forms features such as "a face". This feature is very high-level: it suffices to have two dots and one line, located in a special way relative to each other. And everything else is variable - textures, shapes, angles, colors, gradients, composition, etc.
Another such high-level feature might be something like "composition": the way objects are positioned in a picture, their proportions. Again, this feature doesn't involve shape, texture, color or anything like that.
Yet another such feature might be "shape". There can be sharp or gradual shapes, etc.
Also, color might be such a high-level feature: different combinations of colors, how they relate to one another, how they morph between one another, etc.
We don't know if the models have these exact kinds of features, because it's nearly impossible to peak inside the trained model and extract the features: it's physically possible, but the models are so complex that it's nearly impossible to make sense of the data we see there. But from our understanding of how deep neural networks work, some such features should be there.
These kinds of features are very high-level features, that are definitely not, and cannot be, subject to property rights. Otherwise, humans couldn't lawfully create any art without "stealing" from each other either.
Models infer concrete images from these features. And they can combine the features in ways they weren't combined in the training data. So the models can produce significantly distinct examples from anything in the training data.
Can the models invent completely novel ways of creation, that they didn't see in the training data, examples of epochal significance? For example, could a model invent abstract art having been trained only on representational art? Or could it invent electronic music having been trained only on classical music? Probably, not to the full extent.
But nonetheless, examples these models produce are intricate, complex, rich, detailed, nuanced, intellectually and aesthetically compelling and can be original. It's not necessary that they are always inferior to what humans produce. It can be that a human-produced image is less interesting than a machine-produced one.
Creativity, involved in AI art
One limitation, imposed on generative models is the text input. Not all the richness can be expressed in text, far from it. Text is very high-level way of controlling the output and very limited - very selective.
Because of this lack of control, it could be said that creating art this way is not as creative as creating it by hand from scratch, because in the latter case one has full control over every detail.
On the other hand, creating art by hand from scratch involves a lot of menial, mundane, repetitive work.
It would be very interesting to have more control over generative models. Perhaps, it is possible to extract features a model learned - on various levels of abstraction - and have those features as input to the model. This way one could have much more control, much more fine-grained control as well as high-level control. This way, the need to perform menial work associated with manual creation would be eliminated but fine-grained control would remain. Thus, the creative process would be truly democratized.
Democratization and abundance of art
The cost of producing aesthetically, artistically valuable material with generative AI is orders of magnitude lower than manual production. As a result, the overal quantity of material being produced is orders of magnitude larger. There's simply more aesthetically, artistically valuable material to explore and appreciate.
And this material can occur in places where previously there was no such material. The media space can become much more saturated with good-quality material.
There's also some qualitative differences: for example, generative AI can produce art, tailored to one's specific needs. It can be much more personalized. It becomes much more an on-demand matter.
Previously, if you wanted to have a highly personalized work of art, you had to spend years gaining the skills and knowledge required to produce artworks. But now this highly personalized art becomes accessible to anyone, ordinary people of all walks of life.
This is, without exaggeration, very valuable. And it's very undesirable to prevent this from happening.
I agree that creative process suffers with text-based generative AI, because there's no fine-grained control over all the details of the generated material. But if it's ultimately possible to provide the AI with more detailed input, perhaps in the form of features that it learns from the data, as I described in the previous section, then the creative process too would undergo massive democratization and become abundant.
What to do about "plagiarism"?
There is a significant difference between the machine and a human in the quantity of art each is capable of producing in a unit of time.
Another big difference is that the machine is not sentient, so for it, there's no cost, morally speaking, of producing the quantity of work that it does.
According to current law, producing works that are similar in style to works of other people even in an uncanny way isn't considered plagiat.
But since the machine can produce such work fast and in huge quantities and adapt it's contents to the user's needs, it can easily overshadow the original work to the point where it loses the interest of the consumer. Thus the artist could loose their income.
First of all, it should be said that not all artists will loose their jobs to generative AI. Some of the jobs might be transformed to some degree but a relatively small proportion (7% in US) is estimated to be totally displaced9.
Now, one solution that's suggested to work around the problem of generative AI directly competing with humans is to prevent it from being trained on people's work. But the works, generated by the models are aesthetically, artistically valuable, even if they borrow, in an uncanny way from human-made works. So it's detremental to culture to deplete aesthetically valuable material from it. Beyond that, it might be that the overal quality of the models, their capability to output aesthetically interesting material could significantly suffer if training data is constrained to a relatively small set.
Instead of preventing this material from being generated, it might be sensible to structure the industry in such a way that some of the income from commercial use of these works would go to the artists.
One question here is the ratio between the cost of producing such work and the benefits, reaped from it. On this depends how the industry could be fairly structured in this way.
Another significant question is whether there's an automated way to determine whether a machine-generated work borrows in an uncanny way from an existing work. If there was, then it would be technically trivial to implement a mechanism of compensating artists for machine-generated work in their style.
Another way to deal with the problem of people loosing their income to generative AI is to simply compensate them from the gains from generative AI. I'm not a political or economic expert to judge how fair or feasible this approach is. But it again has the important property that it doesn't prevent the plethora of aesthetically valuable material from being generated.
https://x.com/wesg52/status/1709551516577902782?s=20