Yes, for some texts that's possible. But for the vast majority, it is not.
The New York Times lawsuit is resting on the point that large chunks of undigested articles can be vomited out. OpenAI tried to have the lawsuit thrown out but the courts permitted it to continue.
The Times... alleged that OpenAI's ChatGPT and Microsoft's Copilot had produced near-verbatim replicas of copyrighted articles, that the chatbots generated hallucinated content falsely attributed to the Times, ...
https://en.wikipedia.org/wiki/The_New_York_Times_v._Microsof...
> had produced near-verbatim replicas
Since most users of LLMs are not using them to run around copyright and this copyright stuff is aggressively trained out of models or blocked whenever possible, the models themselves are still transformative works not intended to facilitate infringement.
That's exactly what they've done in a number of the lawsuits, so I'm not sure why you think that hasn't occurred.
https://arxiv.org/abs/2601.02671
The point is that for most texts, it is not possible. It's not able to recall what I wrote on Geocities in 1995, even though there's a good chance it was trained on it.
Or is it simply that the correct prompt hasn't been written for all possible cases?
I also fail to see the difference if logic/harnessing is added around a vector database that can output the complete corpus, but simply is instructed not to.
It very clearly is still compressing the information into the vector weights, and then recovering that information, thus the information is encoded.
Why is a vector database somehow completely different from maintaining a library of the text itself?
LLMs are obviously capable of producing "exact" phrases as well. Ask it to give you famous quotes, it can do it. Ask it to read a paper for you and cite it, it can do it.
I wouldn't bet on that. https://en.wikipedia.org/wiki/Campbell%27s_Soup_Cans
Here's a highly compressed representation of The Lord of The Rings (all three volumes):
1
Obviously, fidelity when uncompressing it is not great, but I can assure you it was lossily compressed from the original text. Is it infringing the original's copyright? I have to assume you'd agree that the answer is "no".
If I had compressed it by removing the letters x y and z, I'd agree with you that my "compressed" version is infringing.
So what we've got here is a spectrum with two ridiculous extremes, and a question: When has the artifact been compressed so heavily that it no longer infringes the copyright of the original?
I suggest "irretrievability" is a pretty good threshold for that question. Otherwise you're into "we know it infringes our copyright. Don't ask us to prove it, we just know it, ok?"
Given the sheer volume of text that an LLM gets trained on, and how small the output is, it seems obvious that 99% of it can no longer be recovered - the process is "lossy" to the point of irretrievability, and only a statistical smear is left behind. That's why I think only the copyright claims that can show infringement in court (Harpy Potter, et al.) have merit. And a court will still have to decide "how much is too much" but at least there's case law for that.
(Incidentally, I compressed the Mona Lisa to a single pixel. It was #3D3526).
I've tried my best to show where I think you're wrong. I think all that's left is arguing over the exact definitions of "recoverable" and "irretrievable". As I said, the courts will have to decide that.
Regardless, the argument that LLM output is or is not subject to copyright based on information theory is entirely defeated by what I've said.
Point an LLM at the conversation and ask it to ELI5 the competing arguments.
> Information entropy. The amount of data an LLM ingests cannot be compressed to the size of the weights even at maximum theoretical compression.
To prove distribution of copyrighted materials it would have to be practical and actually used in the wild by people to circumvent copyright and generate copies of those works. Again, I can't prove a negative, but that isn't the standard, and nobody has shown a practical exploit here.
My assumption is that multiple copies in the training data "wear a deeper groove". I believe those are infringing, and should be dealt with on a case-by-case basis. But the vast majority of text doesn't wear that groove.
(Edit: Think it was this one https://arxiv.org/abs/2601.02671)
That's one hell of a compression ratio, if it can do what you claim.