If you try any modern LLM, you will find that you can. Easily [0], reliably [1], consistently [2]. All these examples are with models released in 2025/26.
[0] https://arxiv.org/html/2601.02671?amp=&=
If its responses were perfect so that you could chain them, or if you could ask "please give me words 10-15 of chapter 3 paragraph 4 of HPatSS, and it did so, then you'd have a better case to complain. Still, the counterargument is that repeated prompting like that, explicitly asking for copyright violation, is the real crime. Are you going to throw someone in prison if they memorize the entirety of HPatSS and recite arbitrary parts of it on demand?
Combining both issues: that LLMs are only regurgitating mostly accurate continuations, and they're only providing that to the person who explicitly asked... any meaningful copyright violation moves downstream. If you record someone reciting HPatSS from memory, and post it on youtube, you are (or should be considered) the real copyright violator, not them.
If you ask for an identifiable short segment of writing, or a piece of art, and get something close enough that violates copyright, that should really be your problem if you redistribute it (whether manually or because you've coded something to allow 3rd parties to submit LLM prompts and feed answers back to them, and they go on to redistribute it).
Blaming LLMs for "copyright violation" is like persuading a retarded person to do something illegal and then blaming them for it.
What is the real copyright risk of there being an arcane procedure to sometimes recover most of a text? So far it’s nothing. Which is what I’m saying. Pragmatically this is a loser of an argument in a court room. It is too easy for the chain of reasoning to be disrupted and even undisrupted the argument for model maker liability is attenuated.
I have, on many occasions, gotten an LLM to do just this. It's not particularly hard. In the most recent case google's search bar LLM happily regurgitated a digital ocean article as if it was it's own output. Searching for some strings in the comments located the original page and it was a 95% match between origin and output.
> The memorization older models showed came from problems in the training data,
And what proof do you have that they "fixed" this? And what was the fix?
> harry potter and people writing about harry potter
I'm not sure that's how you get GPT to reproduce upwards of 85% of Harry Potter novels.
> Second, infringement would need discovery to uncover and would be contingent on user input.
That's not at all how copyright infringement works. That would be if you wanted to prove malice and get triple damages. Copyright infringement is an exceptionally simple violation of the law. You either copied, or you did not.
> For copyright it very much matters if something is lifted verbatim vs modified.
Transformation is a valid defense for _some_ uses. It is not for commercial uses. Using LLM generated code for commercial purposes is a hazard.
We have yet to see a single judgment come down against a model maker for distributing the gist of content. We have yet to see a single judgment come down against a model maker for infringement at all.
Copyright is just an inapt tool here. It’s not going to do the job. It is not as though big interests have not tried to use this tool. It just doesn’t reflect what’s actually happening and it’s going to lose again and again.
We can imagine a theoretical legal regime where what is done with large language models counts as copyright infringement, we just don’t live in a world where that regime holds.