The New York Times is suing OpenAI and Microsoft for copyright infringement, claiming the two companies built their AI models by “copying and using millions” of the publication’s articles and now “directly compete” with its content as a result.
As outlined in the lawsuit, the Times alleges OpenAI and Microsoft’s large language models (LLMs), which power ChatGPT and Copilot, “can generate output that recites Times content verbatim, closely summarizes it, and mimics its expressive style.” This “undermine[s] and damage[s]” the Times’ relationship with readers, the outlet alleges, while also depriving it of “subscription, licensing, advertising, and affiliate revenue.”
The complaint also argues that these AI models “threaten high-quality journalism” by hurting the ability of news outlets to protect and monetize content. “Through Microsoft’s Bing Chat (recently rebranded as “Copilot”) and OpenAI’s ChatGPT, Defendants seek to free-ride on The Times’s massive investment in its journalism by using it to build substitutive products without permission or payment,” the lawsuit states.
The full text of the lawsuit can be found here
No it doesn’t, the training data isn’t inside the LLM.
So firstly, even if those claims are true, you sue the wrong business, you would need to sue the training data maker. They however are usually protected by laws for science, because they are “non profit research”
Therefore this is completely ridiculous.
Btw, A the copyright part is only a thing if its a significant portion of the thing… Wich it clearly isn’t in this case (its below 1% of it) making it even more ridiculous.
Also, if you can get the information on the internet, you are again suing the wrong place, you should be after the provider, not the automatic data grabbing system… As they can and will argue that they cant control what their algorithm crawler takes. There is a way to mark content as “dont use” for Mashines, but most people don’t do that and will lose in court because they don’t understand it…
Lastly, the training wouldn’t be harder, the problem is the gathering of data. You can’t manually look through all of it and its idiotic to think that its reasonable to demand such a thing.
This is factually incorrect. You can extract the data. How do you think the legal cases are being brought?
For example
The model has to contain the data in order to produce works.
Wholesale commercial copyright infringement where you’re profiting off of others work on a large scale is a whole different ball game.
They’re training their models on large amounts of pirated content and profiting off it.
Of course the rights holders are going to say “wait a minute, why are you making money off my content without my permission? And how much of my work did you pirate to use?”
You cannot hand wave away mass piracy to train their models, and then distribute said models based on an act of mass copyright infringement.
Do you not understand the basics of the law?
Again, the law is the law. If they mass pirate a bunch of media which then the model contains chunks of they are breaking the law.
I can’t believe this is a hard concept for someone to understand.
Even if the compression is extremely lossy, compression is insufficient to be transformative.
The whole “the original data isn’t in the model” argument is one only techbro idiots find compelling.
No, that’s the current legal precedent within the US.
Kelly v. Arriba Soft
The court opinion:
That “compression is transformative” principle has been pretty solidly enshrined as precedence at this point (IE Perfect 10, Inc. v. Amazon.com, Inc.) however with no real guidelines as to what amount is required to be considered transformative
The major argument as to whether the sort of LLM training in the parent article still constitutes fair use or not depends on whether there exists “market harm” or the “substantiality of copying” is especially egregious (note that these are the two fronts that the NYT is taking.) There is precedence for copying of style not being fair use Dr. Seuss Enters., L.P. v. Penguin Books USA, Inc. which I suspect is why NYT is approaching it the way that they are…
Now, all that being said, my personal opinion is fuck the US legal system and fuck copyright. There is no solution to the core issues surrounding this topic that isn’t inherently contradictory and/or just a corporate power grab. However, the “techbro idiots” are “right” and you’re not, but it’s because they are idiots who are largely detached from any sort of material reality and see no problem with subjecting the rest of us to their insanity.
Some form of copyright has to exist, and - as angrily explained to me by authors - it needs to extend somewhat beyond the life of the author. I’m certainly never going to agree with it being indefinite though.
I can’t tell if sarcasm… If not why?
Authors need to be able to make a living from writing, unless you want far less books to be written. And there is some logic in allowing them to leave some sort of rights to their family, even if it’s only for like 10 years.
deleted by creator
The model has to contain the data in order to produce works.as far as I understand, this isn’t true. can you elaborate on why it needs to contain the data?
It contains large parts of the data in order to create. In my link I provided it shows that the models do contain chunks of the original works.
Otherwise, how would it create the words etc.
I am amazed that we now have people on the level of crypto coin idiocy going on about ai models who don’t understand this.
You would probably claim I don’t deserve my job with my level of technical illiteracy however you think you are inferring that . Anyways they do make reasonable efforts to design models that don’t memorize and are able to generalize. This is quite basic or fundamental on machine learning in general.
Previous models had semantic reasoning capacidad without memorization e.g. word2vec.
You should also realize that just because current models are memorizing despite efforts to prevent it doesn’t mean that models need to memorize. Like i said initially they are actually designed to work without needing to memorize.
You’re contradicting yourself.
In one sentence you say it doesn’t memorize (with “reasonable effort”) then in the next you admit it does.
“Reasonable effort” is weasel wording.
Make up your mind.
?? Are you trolling. If you design a car to combust gasoline without burning the lubricants but you still end up burning them it doesn’t mean that the lubricants are needed for the combustion itself. Conversely you have not made any nuanced argument explaining why memorization is necessary. I gave you an example where we know there is no memorization and you ignored it.
“Otherwise how would it create the words” is just saying you wouldn’t know.
?? Are you trolling. If you design a car to combust gasoline without burning the lubricants but you still end up burning them it doesn’t mean that the lubricants are needed for the combustion itself. Conversely you have not made any nuanced argument explaining why memorization is necessary. I gave you an example where we know there is no memorization and you ignored it.
“Otherwise how would it create the words” is just saying you wouldn’t know.
So, me pointing out the flaw in your argument is trolling?
What?
If you choose to use weasel wording to try and get out of something that is your call.
Ok i believe that you believe that. It’s ok. I have professional experience in this space so you’re either not reading carefully or you don’t understand much about the topic.
Perhaps you might want to reconsider this in more abstract terms. The engine example you ignored could help you with that.
Do you really think that the fact that we have language models that don’t memorize and are simple enough that we can know for certain is not all we need to show that language models don’t necessarily have to memorize? You keep repeating the same (illogical) argument and ignore the simpler arguments that disprove your claim.
This entire comment screams of 0 technical knowledge.
The LLM does not contain the training data. It contains nothing but math it generates you an answer by calculations, in the end you get the awnser wich is statistically most likely what you want. Otherwise the fucking thing wouldn’t produce fake news and make shit up.
Shure if you want it to write you a very specific thing and you know exactly what to ask, you might get a small text that is “copyrighted” but thats because you asked for it, not because it’s inside. It just gives you the awnser you most likely find helpful, statistically.
Its like asking you to read a page very well and then asking you the next day to write down what was on the page, while giving you lots of hints. You didn’t actually copy from it in that case.
My guy, if you compellingly re-wrote Harry Potter from memory and charged people for access to your work, you can definitely expect J.K. Rowling to sue you.
Yes, your comment does.
There is literally software to extract this stuff from models now.
This “it’s just math” is techbro idiocy. It’s like the idiots regurgitating crypto coin bullshit.
It’s all black magic to me, so if you have resources on this, that would be great. My initial thought is that it would have surely have a data source to reference to? Your last example is some one referring to their memory of something and recreating it. By referring to that memory, that is in essence a reference back to the original data that someone has remembered?
the poem poem poem thing shows that the llms actually do memorize at least some training data. chatgpt changed their eula to forbid users from asking it to repeat words forever after this was in the news.
also as far as I understand there are usually fair use and non profit exceptions for use of training data but they generally limit how it can be used. so training a model for commercial purposes might be against the license of the training data.
I don’t necessarily agree with the nyt but they seem to be framing this as someone aggregating their data and packeting it in a better way so they are hurting their profits. i don’t really see that as necessarily being true. they could argue the same about google news showing their news…
They don’t “remember” anything they produce a “awnser” by generating a shit load of math wich renders down to the most “helpful” answer it can statistically give you.
LLMs are neuronal networks, if you know how they work you know how idiotic all copyright claims are, they all just mad that their shit is getting obsolete and in the background use the engine to do “work” wich they claim to have violated their copyright, now they are mad because it does a better job at writing than they do and they fear of being replaced.
All lawsuits against AI companies, regarding copyright of training data, are dumb as hell.
You are right about the commercial/non profit training data part, but from my understanding that’s basically a gray zone and politics are to slow to keep up with tech.
Btw fuck Open AI, they are as open as a fucking Supermax prison. Even the programmers don’t know what their main LLM does, they just place a simple one between the user and the actual GPT to make shure that it doesn’t give people instructions on how to build a bomb and stuff like that or to keep people from making it say bad words…
that’s the theory. previous models also were supposed to be doing 3 digit math but they dicovered that the questions were in the training data.
so you should look into what happens when people ask chat gpt to repeat a word forever, it prints the word for a while and then prints training data, check this link https://www.404media.co/google-researchers-attack-convinces-chatgpt-to-reveal-its-training-data/
edit: relevant part:
I should also reiterate that I agree that the intent is to avoid memorization, but they are not successful yet.