You won't find a copy of any specific work anywhere in the compressed form of a file, either, but when you decompress it you find the complete work. And many large AIs can recite, verbatim or near-verbatim, many complete works. Yes, they might get a word wrong, but that doesn't nullify the point that they're trained on the entire work and to a first approximation they can emit the whole work.
> Typically the consumers of the weights are software developers or content creators, whereas the consumers of the original text or image are fans.
Many of the consumers of image models are in fact generating art that they previously would have commissioned from an artist. (Some of them are also generating art they never would have commissioned from an artist, so I'm not implying that this is a one-for-one revenue loss.) Consumers of code models are, in fact, potentially reducing the total demand for novice programmers.
The model derived from a pile of artistic works is, in fact, directly competing with those artistic works.
Sure you will. It's right there, in PNG encoding or what have you. With nothing more than the compressed file and general purpose tools you can reliably put it on your screen.
> And many large AIs can recite, verbatim or near-verbatim, many complete works.
This is not the common case and it's not even clear that the reason it can sometimes do this is having been trained on the complete work. The more likely cause is having seen a large number of excerpts from the work which patch together into the whole thing, because it's more likely to output that text if it has seen it multiple times.
This is also clearly not the intent, purpose or typical use of the model. It has no ability to consistently do that.
> Many of the consumers of image models are in fact generating art that they previously would have commissioned from an artist. ... Consumers of code models are, in fact, potentially reducing the total demand for novice programmers.
Those are different people than the holder of the copyright on an arbitrary piece of the training data. Is a fair use determination supposed care about the effect on the market for some entirely independent work from an unrelated third party?
> The model derived from a pile of artistic works is, in fact, directly competing with those artistic works.
It's indirectly competing with artistic works in general.
I mean take the computer out of the loop and think about it this way. Art teachers everywhere reproduce a bunch of existing works for classroom use to train new artists who go into competition with the original artists. That is obviously going to indirectly impact the market for artistic works in general, but I don't think that's how that factor works.
No, I'm saying many AI models use copyrighted works by independent artists to directly compete with those artists (in addition to other artists). And AI models use copyrighted works by software developers to compete with those software developers (in addition to other developers). Fair use determinations do care if you're competing with the work you copied.
> The purpose and character is absolutely heavily commercial and makes a great deal of money for the companies building the AIs.
That's assuming that it's Microsoft/OpenAI. Suppose a non-profit trains a model and releases the weights for free.
> There's nothing about the works used for AI training that makes them any less entitled to copyright protections or more permissive of fair use than anything else. They're not unpublished, they're not merely collections of facts or ideas, etc.
The models aren't trained on works of a particular nature, they're trained on whatever they can find, so this doesn't really mean anything until you're talking about a specific work.
> AI training uses entire works, not excerpts.
The weights don't contain entire works. They contain statistics about entire works, but that's not the same thing. You can't find a copy of any specific work anywhere in the weights. Nobody can give you a piece of code that will decode the weights into all the original works.
> AI models are having a massive effect on the market for and value of the works they train on, as is being widely discussed in multiple industries.
That's not how that factor works (and it's the most important one). No one is buying the model weights so they can read them like a novel. Typically the consumers of the weights are software developers or content creators, whereas the consumers of the original text or image are fans.
To make this a little clearer, suppose the purpose of the model isn't to generate content, it's to generate recommendations. Then the company operating it takes a list of content anyone likes, uses the model to show you a list of all the other content you might like and lets you sort by price. Which makes it easier to find competing content which is available for less money, which increases competition. The incumbents might hate this, and it might even lower their profits, but that's not the kind of effect on the market this factor is supposed to be about.
Moreover, suppose you had a model trained entirely on public domain works. Obviously this can't be copyright infringement even if it's extremely effective at producing new works that compete heavily with works still under copyright. But if you added some specific work still under copyright to the model, it would only be a marginal difference. The effect on the market for that specific work of adding that specific work to the model would be negligible. It's the technology itself that provides the competition, not the accretion of any particular work in the weights.