LLMs - as in the models, the weights - do not contain "reproduction of fictional works or the creation of derivative works based on fictional training inputs" in any meaningful sense. "Reproduction of fictional works" is what is present in the training data. "Creation of derivative works based on fictional training inputs" is what you might be doing as a user of a LLM. The models themselves are, like GP said, recordings of statistical properties of text, just taken up to 11.

> do not contain "reproduction of fictional works or the creation of derivative works based on fictional training inputs" in any meaningful sense.

Except it does, and it's in a quite literal sense.

You painted a painting. I took a picture of it and compress it as .jpeg. Byte-wise the .jpeg file has no similarity to your original painting. For someone who doesn't know .jpeg, it's just garbage bytes.

So I didn't infringe your copyright by selling this .jpeg file. The user who decodes the .jpeg file and displays it on a monitor does.

Does it sound right? This is how "weights do not contain training data" argument works.

And before "how about artists who store this information in their brains?" Well human being is a special case for every law in every country. Just like selling a cow's liver is never the same as selling a man's, even they're both organic tissues. A human's brain is always going to be treated differently than a hard drive.

My argument is that the modes are somewhere between your JPEG example and "artists' brains" example. Now legally, it's usually not the bytes themselves that matter, but their colour (provenance). But this doesn't make the case of AI models any more clear - the training process is also somewhere between zipping up a folder full of JPEGs, and a human practicing their art through fuck ton of repetitive reproductions of existing works until they grok the style.

I agree.

The thing I hate about this whole situation is that it's going to be decided solely on which side has more lobbying power[1], since just as you said, it's something in between.

[1]: One could say all the policies are decided this way...

> [1]: One could say all the policies are decided this way...

The silver lining here is that lobbyists can only afford to care about finite number of things at a time - so even if they get some laws their way, other factions can push laws mitigating the damage somewhat.

I currently believe (weakly) that the outcome of more and better lobbying isn't laws getting directly worse for society, but rather the regulatory system grinding to a halt under increasing number of laws that exist to cancel out parts of other laws...

> LLMs - as in the models, the weights - do not contain "reproduction of fictional works or the creation of derivative works based on fictional training inputs" in any meaningful sense.

Since portions of such works can be recovered in inference, they contain at least a lossily-compressed copy of the collection of works used in training. Reproduction isn’t pure coincidence.

Indeed, but learning is effectively[0] a form of lossy compression too. DNN weights are somewhere between zipping up a truckload of JPEGs and deriving facts from first principles. Where exactly are they on this spectrum and how it affects copyright, it's not obvious to me - at least not when trying to argue from fundamental principles. I feel the DNN copyright issues will be ruled on from purely pragmatic position: what's the legal status that upends existing markets the least, and/or is most favored by lobbyists.

--

[0] - I believe it's actually fundamentally the same thing.

> Indeed, but learning is effectively a form of lossy compression

Yeah, that copies in the brain of a human who has experienced the work and not reduced to any othe media are neither “copies” nor “phonorecords” as covered by US copyright law is pretty clear, and that this does not apply to data stored by clever algorithms in computer storage is also pretty clear, so I’m not sure what your point is.

If you want to argue LLMs are people to escape this, then, sure, copyright stops being a problem for their training data, but it still is for their output plus you end up a whole set of new legal problems with using them as people do now, starting with 13th Amendment problems.

There's a pretty big jump from "machine learning is not inherently different from human learning" to "machine learning models are people".

There's not a big jump from “human learning isn't treated as a copyright violation because it happens in people and not in external media” plus “machine learning should be legally treated like human learning” to “your argument depends on treating ML models as people”.

The fact that people making the second part of the first argument often either are ignorant of or deliberately ignore why human learning is treated the way it is doesn't change the essence of the situation.

That applies to both sides though: people arguing that machine learning should not be treated like human learning often justify it with "ML models aren't people, they're software on a computer". But that argument is also missing the point, and invites this kind of philosophical discussion. The essence of the matter is that there is no high-level principle involved - humans get special treatment because humans write the law. So we can legally round all machine learning down to "zipping a folder full of JPEGs" for pragmatic reasons, and get away with it until we actually create sentient AIs and recognize them as people.

(Which, knowing history and observing how we treat animals, will happen only when the sentient AIs coerce us to recognize them, through violence or the threat of it...)

> The essence of the matter is that there is no high-level principle involved - humans get special treatment because humans write the law.

“The purpose of law is to advance the collective interests of humans in a society” is, I would argue, a high-level principal, and (with variations as to whether it applies to all humans or some subset, and which interests are considered privileged relative to others) nearly universal. Yes, the fact that humans are writing the laws is a reason that principle is chosen, but the idea that there are not high-level principle is just false.

>“your argument depends on treating ML models as people”

They would be "legal people" in the sense that corporations are "people".

> that this does not apply to data stored by clever algorithms in computer storage is also pretty clear,

I'm sorry, this is far from "clear", otherwise we wouldn't be having this discussion. Fair use is a thing. Does fair use apply in this situation? No one knows at this point.

I think you're conflating two issues. As I understand it, this justification is about whether the training itself is fair-use. The discussion on whether the outputs are derivative or transformative is separate. Analogously, if I publish a copy of copyrighted poems from memory that would be infringement, but there's nothing infringing about me just reading and memorizing them.