Comedian and author Sarah Silverman, as well as authors Christopher Golden and Richard Kadrey — are suing OpenAI and Meta each in a US District Court over dual claims of copyright infringement.
Interested to see how this plays out! Their argument that the only way a LLM could summarize their book is by ingesting the full copyrighted work seems a bit suspect, as it could’ve ingested plenty of reviews and summaries written by humans and combined that information.
I’m not confident that they’ll be able to prove OpenAI or Meta infringed copyright, just as i’m not confident they’ll be able to prove that they didn’t violate copyright. I don’t know if anyone really knows what these things are trained on.
We got to where we are now with fair use in search and online commentary because of a ton of lawsuits setting precedent, not surprising we’ll have to do the same with machine learning.
ThePile, which was assembled by a company called EleutherAI. ThePile, the complaint points out, was described in an EleutherAI paper as being put together from “a copy of the contents of the Bibliotik private tracker.” Bibliotik and the other “shadow libraries” listed, says the lawsuit, are “flagrantly illegal.”
I think this is where the crux of the case lies since the article mentions these are only available illegally through torrents.
This is starting to touch on the root of why they keep calling this “AI”, “training”, etc. They aren’t doing this for strictly marketing, they are attempting to skew public opinion. These companies know intimately how to do that.
They’re going to argue that if torrents are legal for educational purposes (ie the loophole that all trackers use), and they’re just “training” an “AI” then they’re just engaging in education. And an ignorant public might buy it.
These kinds of cases will be viewed as landmark cases in the future and honestly I don’t have huge hopes. The history of these companies is engineer first, excuse the lack of ethics later. Or the philosophy of “it’s easier to apologize than ask”.
It’s the defacto term for how we fit a statistical model to data, unrelated to any copyright concepts. I’m pretty sure we called it “training” back in 1997 when I was doing neural networks at uni, and it’s probably been used well before then too.
Neural nets are based on the concept of Hebbian learning (from the 1930s), because they are trying to mimic how a biological neural network learns.
This concept of training/learning has persisted because it’s a good analogy of what we are trying to do with these statistical models, even if they aren’t strictly neural networks.
Even if they did train the model on the entire text of the book, that’s still not necessarily copyright violation. I would think not, since the resulting model doesn’t actually have a copy of the book embedded within it.
It’s difficult to tell to what extent books are encoded into the model. The data might be there in some abstract form or another.
During training it is kind of instructed to plagiarize the text it’s given. The instruction is basically “guess the next word of this unfinished excerpt”. It probably won’t memorize all input it’s given, but there’s a nonzero chance it manages to memorize some significant excerpts.
It’s difficult to tell to what extent books are encoded into the model. The data might be there in some abstract form or another.
This is a court case so the accusers are going to have to prove it.
The evidence provided is that ChatGPT can produce two-page summaries of the books. The summaries are of unknown accuracy, I haven’t read the books myself so I have no idea how much of those summaries are hallucinations. This is very weak.
But the server used to calculate the model would have a copy of it. If training an AI model is not fair use then the mere act of loading a book you don’t have a license for into the server would be copyright infringement. Like text book. It’s a unauthorized digital copy. It’s all very untested legal grounds and seems like lots of people want to be the first to test it. Not everyone has a great case but if the courts interpret things a certain way there’s gonna be lots of payouts so maybe best to get in line early?
Perhaps, but that’s a separate legal issue from the model itself. You might have committed a breach of copyright in the process of gathering the material that the AI was trained on but the model itself is not a copy of that material and so is not itself illegal to train or use. And perhaps not even that, since downloading a pirated book is not the illegal part (uploading it is).
As you say, there’s some untested legal waters here. But it seems likely to me that the best that Silverman will accomplish is some nibbling and quibbling around the edges.
How do we “know” anything where the answers are just being made up as part of humanity’s collective cultural game of Calvinball?
Courts in various jurisdictions will make various rulings. Judges will interpret them in various ways. Legislators will chime in with new legislation and new treaties. Internet arguments will churn away with a whole range of assumptions about what is true or false that may or may not have anything to do with reality.
I present my opinion here. I feel it is well informed and I can back it up in various ways when challenged. But nobody “knows” anything because these aren’t laws of physics or math that we’re talking about here.
Or did you mean whether we know if a copy of the book is embedded in the model? That can be more objectively tested, at least.
AFAIK it takes these large bodies of text and rather than digesting them and keeping it in some sort of database, rather it holistically (and i’m generalising here), see how often certain words are strung together and taking note of that. Let’s call them weights.
Then users can prompt something and the ‘magic’ here is that it is able to pick out words of different weights based on the prompt. Be it, are you writing an angry email to your boss, a code in python, or structure for a book.
But it is unable to recreate the book from a prompt.
People who know the topic more intimately please correct me if I am wrong .
It may be that no one currently knows exactly what these things are trained on, but it could be determined. If you know the methodology you can figure out what data is being used. The companies involved are going to resist letting anyone find out, but I’m hoping a court case will break that black box open.
One of the many problems with this form of AI is the degree to which we don’t know where it’s getting its information from. Without that, there is no way to determine the reliability of the results. They can sound perfectly reasonable and be entirely untrue.