Slop Is Plagiarism | Hacker News Reader

Slop Is Plagiarism

techrights.org

3 pointsamcclure10 comments

'Plagiarism' is an overreach. An LLM works a lot like the human brain does in terms of how it gathers its ideas for what to output, just like people do. When you write or create art, your ideas are a synthesis of everything you've experienced externally, directed and focused by the context of what it is you are looking to create/output. Nothing in your head is more original than what an LLM might produce. If you hadn't listened to music composed by other composers for most of your life you wouldn't be able to generate a new composition and neither would an LLM. Are you plagiarizing when you write a new composition because you listened to famous composers and practically memorized their music over your training run (your life)? I'd say no. Plagiarism is more like the intentional act of copying work instead of doing your own synthesis.

Is a painter who was trained on Van Gogh or Salvador Dali plagiarising those artists when their own work shows influences of their work? Meh, I dont think so

Is Greta Van Fleet plagiarising Led Zepplin? Some would say so out of distaste but it's not an actual fact.

You actually have no idea if an LLM works "a lot like the human brain". No one does. It is still not clear if neural networks are capable of generalization to data outside of the training set in any capacity. Not just LLMs.

If they could, why would we need these monstrous models, even for simple problems?

I didn't say that it works like a human brain mechanistically, I said that in terms of how it can only output variations of the underlying training data, it does, indeed resemble (or 'works like', in common parlance) how a brain works. Your confusion stems from your literal take on the term works like.

The theory is sound.

A lot of artists are doing mix and match and they're fine because it's framed as artistic interpretation. Can the author point to examples of LLMs reproducing some original work without being explicitly asked to do so? If not, what's the argument?

I read the source code and like some patterns that I reuse in my later works. Is that plagiarism? If not, what's the difference with LLMs in this sense?

I for one am still baffled at just how OpenAI/Anthropic and the likes have gotten away with the biggest copyright infringements (or perhaps IP-theft, I'm not sure what the correct term is) in history (the data collection for training) without any legal ramifications.

Even if it turns out to be perfectly legal I think a "retreat" to this position of LLM generation being Plagiarism (and therefore being uninteresting and morally bad) is quite a reasonable action.

It also separates LLMs for searching through documents quite nicely from content generation which since I would be hard pressed to consider a "fuzzy search" as a form of plagiarism.

If they buy the books, is it still theft?

For training purposes? I'd say no.

The legal or ethical status of generated output is still up for discussion. I myself do not think neural networks can generalize and thus I would have to say that it still would be copyright infringement since its just regurgitation of input data. Just because I bought a book doesn't mean I get to re-publish it.

they are not outputting or "regurgitating" like that, especially as the models get smaller, they just cannot encode that much information for reproduction, not so dissimilar from humans

maybe a couple of years back in the early experiments, but not today, if it were easy to get them to do it still, the activists would have by now

Yes they are. Just because the models have been overfit on the entirety of the internet doesn't mean they aren't in principle re-generating parts of their input data for the overwhelming majority of their output. Sure, stochasticity pushes the output into weird states sometimes, these are the states where the output is plain useless nonsense. Hallucinations, if you will.

Our brains are entirely too small to recognize the patterns encoded from the gigantic training data set. But they are there nonetheless.

This is easily seen if you try get *any* neural network to generalize to *any* non-trivial task (in this case, XOR for example I consider to be trivial). There are no non-flickering, stable vision models that can recognize a human/dog/cat/whatever in all the lighting settings we can without false positives. The models are only getting better because they are getting bigger and are provided with more data. This is *not* generalization.

> Yes they are.

If you change the definition to be more flexible, sure, you can claim such things. But I will disagree because under this definition humans are largely just regurgitating too. Consider how evolution has been a similar training regime for organic systems.

Reply on news.ycombinator.com