OpenAI’s ChatGPT and Sam Altman are in massive trouble. OpenAI is getting sued in the US for illegally using content from the internet to train their LLM or large language models
This is interesting. Now wealthy folks can defend their copying of data for personal gain while the concept of content piracy is a criminal offense for the everyday joe, complete with steep fines and sometimes vacations to clubfed.
I very much enjoy ChatGPT and I’m excited to see where that technology goes, but lawsuits like this feel so shaky to me. OpenAI used publicly available data to train their AI model. If I wanted to get better at writing, and I went out and read a ton of posted text and articles to learn, would I need to go ask permission from each person who posted that information? What if I used what I learn to make a style similar to how a famous journalist writes, then got a job and made money from the knowledge I gained?
The thing that makes these types of lawsuits have a hard time succeeding is proving that they “Stole” data and used it directly. But my understanding of learning models in language and art is that they learn from it more so then use the material directly. I got access to midjourney last year August, and my first thought was, better enjoy this before it gets sued into uselessness. The problem is, people can sue these companies, but this genie can’t be put back into the bottle. Even if OpenAI get hobbled in what they can do, other companies in other countries will do the same and these law suits will stop nothing.
We’re going to see this technology mature and get baked into literally every aspect of life.
Absolutely agree with you. It’s in theory no different to a child learning from what they’re exposed to in the world around them. But I guess the true desire from some would be to get royalty payments every time a brain made use of their “intellectual property” so I don’t think this argument would necessarily convince.
I think the more realistic way it would be handled is that the site you published your content on puts in its terms of service that you agree that stuff you put on their site can be used for AI training, and openAI buys the data from them, either via an API key or a data dump or whatever.
But I see merit in not allowing companies to profit off of whatever content already exists without any sort of consent. And I don’t agree with the idea it’s like a child learning… You aren’t raising a child to sell a subscription to their knowledge and profit off of it
if you release data into the public domain (aka, if it’s indexable by a search engine) then copying that data isnt stealing - it cant be, the data was already public in the first place.
this is just some lawyer trying to make a name for themselves
Just because the data is “public” doesn’t mean it was intended to be used in this manner. Some of the data was even explicitly protected by gpl licensing or similar.
but GPL licensing indicates that “If code was put in the public domain by its developer, it is in the public domain no matter where it has been” - so, likewise for data. if anyone has a case against OpenAI, it’d be whatever platforms they scraped - and ultimately those platforms would open their own, individual lawsuits.
That’s not at all how the GPL works…
can you expand on that? I’m not very familiar with the legal aspect of GPL.
crickets…
If you release code under gpl, and I modify it, I’m required to release those modifications publicly under gpl as well.
so if content is under GPL and used for training data, how far is the process of training/fine-tuning considered “modification”? For example, if I scrape a bunch of blog posts and just try to use tools to analyze the language, does that considered “modification”? What is the minimum solution that OpenAI should do (or should have done) here, does it stop at making the code for processing the data public, or the entire code base?
I’m not sure. And I’m not sure there’s legal precedant for that either.
That’s why I dont have a problem with any of these lawsuits, it gives us clarity on the legal aspects, whichever way it goes.
Not a lawyer, but you can argue that if the language model is trained using gpl licensed data, then the language model has to be published under gpl as well.
Not everything indexed by a search engine is public domain that’s not how copyright works.
There’s plenty that actually is in the public domain but I guess scraping the web is a lot easier for these people
I don’t agree. Purpose and use case should be a factor. For example, my friends take pictures of me and put them on social media to share memories. Those images have since been scraped by companies like Clearview AI providing reverse face search to governments and law enforcement. I did not consent to or agree to that use when my likeness was captured in a casual setting like a birthday party.
perhaps - but it could easily be argued that you knew that what you share on the internet was viewable by anyone. are you going to sue Clearview and/or the law enforcement agencies for control over your image that’s in the public domain?
Public Domain
You keep using that word. Maybe you should look up what it means.
You know that you are making a photo public when you post it publicly
I didn’t post it
Let’s note that a NY Magazine article is copyrighted but publicly available.
If an LLM scrapes that article, then regurgitates pieces of it verbatim in response to prompts, without quoting or parodying, that is clearly a violation of NY Mag’s copyright.
If an LLM merely consumes the content and uses it to infinitesimally improve its ability to guess the next word that fits into a reply to a prompt, without a series of next-words reproducing multiple sentences from the NY Mag article, then that should be perfectly fine.
That’s not how copyright works. You cannot freely monetize on other people’s work. If you publish some artwork I cannot copy it and sell it as my own work.
But you can learn from it and create your own new art that may have a similar style as the original
A human can, within limits.
But software isn’t human. AI models aren’t “learning”, “practicing” and “developing their own skills”.
Human-made software is copying other peoples work, transforming it, letting a bunch of calculations loose on it, and mass producing similar works as the input.
Using an artists work to train an ai model and making similar stuff with it to make money off of it, is like copying someones work, putting on a mug, and selling that.
It’s not using it as inspiration to improve your own skills.People are humanizing computer programs way too much, and thus, we have arguments like this. An AI language model is not one of those sci-fi AIs that live in spaceships and talk to the crew. AI language models do not have individuality, creativity, consciousness, or free will. They are computer programs doing math to turn inputs into outputs.
deleted by creator
I don’t agree. Purpose and use case should be a factor. For example, my friends take pictures of me and put them on social media to share memories. Those images have since been scraped by companies like Clearview AI providing reverse face search to governments and law enforcement. I did not consent to or agree to that use when my likeness was captured in a casual setting like a birthday party.
If I learned to read from Dr. Seuss books, does that mean that everything I write owes a copyright tariff to the Geisel estate?
No. But what if you spoke only by reproducing small fragments of text verbatim from The Lorax?
So anyone who creates something remotely similar to something online is plagiarizing, got it.
Folks, that’s how we all do things - we read stuff, we observe conversations, we look at art, we listen to music, and what we create is a synthesis of our experiences.
Yes, it is possible for AI to plagiarize, but that needs to be evaluated on a case by case basis, just as it is for humans.
The lawsuit isn’t about plagiarism; it’s about using content without obtaining permission.
And exactly which AI is republishing content unmodified?
We are creating content based on this article, but no one is accusing us of stealing content. AIs creating original content based on their “experience” is only plagiarism (or copyright violation) if it isn’t substantially original.
And exactly which AI is republishing content unmodified?
We are creating content based on this article, but no one is accusing us of stealing content. AIs creating original content based on their “experience” is only plagiarism (or copyright violation) if it isn’t substantially original.
Is it stealing to learn how to draw by referencing other artists online? That’s how these training algorithms work.
I agree that we need to keep this technology from widening the wealth gap, but these lawsuits seem to fundamentally misunderstand how training an AI model works.
algorithm scraping the web for data doesnt play by the same rules as humans mate
AI is not human. It doesn’t learn like a human. It mathematically uses what it’s seen before to statistically find what comes next.
AI isn’t learning, it’s just regurgitating the content it was fed in different ways
But is the output original? That’s the real question here. If humans are allowed to learn from information publicly available, why can’t AI?
No, it isn’t original. Output of AI is just reorganized content that it already has seen.
AI doesn’t learn, it doesn’t create derivative works. It’s nothing more than reshuffling what it’s already seen, to the point that it will frequently use phrases pulled directly from training data.
You are saying that it isn’t original content because AI can’t be original. I’m saying if the content isn’t distinguishable from original content, and can’t be directly traced to the source, in what way is it not original?
Because it’s still not creating anything. AI can’t create, it just reorganizes.
I think you hear a lot of college students say the same thing about their original work.
What I need to see is output from an AI, and the original content side by side and say “yeah, the AI ripped this off”. If you can’t do that, then the AI is effectively emulating human learning.
I think you hear a lot of college students say the same thing about their original work.
What I need to see is output from an AI, and the original content side by side and say “yeah, the AI ripped this off”. If you can’t do that, then the AI is effectively emulating human learning.
No it isn’t
AI is math. That’s it. This over humanization is crazy scary that people can’t see the difference. It does not learn like a human.
It’s a well established problem. Tech companies have explicitly told employees to not use these services on company hardware or servers. The data is not abstracted from the user and it’s been proven to output data that’s been inputted.
There’s a line here that is a little ambiguous.
If I create a program that’s designed to learn to play video games, do I need to specifically get the consent of the developers of all games that I have legal access to? Do I need to be able to redistribute a piece of IP before I can make use of it to train an AI?
That doesn’t seem right.
Do I need to own a copyright before I can use Dark Reader on a webpage? To use accessibility software? Ad blockers?
Do I need to own a piece of music in order to learn to play it? To learn about composing from it and take it as a source of inspiration?
It seems to me that if you’re putting your content out there for all the world to see, the world seeing that through the lens of a program they wrote and making use of that experience to teach their program to understand language and visual representations ought to be within the realm of the reasonable and expected.
We live in a world where our data is gathered sneakily on a regular basis in order to build massively invasive personality profiles on us that do us no good and make a massive profit for others. Everybody’s data is already being stolen. But this uses information that’s out there for anyone to take and hands us something of incredible value in return that gives tremendous power to individuals. It learns from us and we learn from it. Seems like a fair trade.
LLMs are a tremendous resource that we really need to protect public access to.
It’s not a fair trade if there is no consent.
The main difference is that if you as a person learn music, or play video games, or anything else where you take someone else’s work and make it your own, you are making it your own.
ChatGPT and other AI like it only regurgitate their training data in a mishmash of almost, but not quite, nonsense. There’s no “reimagining” there’s no creativity, it’s just a literal rehash of the training data.
It’s even in the name, the P in ChatGPT stands for “pre-trained”. It isn’t learning anything new, it’s just spitting out bits and pieces of what you originally fed it, and that is copyright infringement with extra steps.
About the only two major places on the Internet that explicitly give permission to reuse their users’ content are Wikipedia and Stack Overflow. Add the public domain texts in Project Gutenberg and that’s about it, the rest of the Social Media requires consent from the user to be reused elsewhere. This is the crux of the problem, CEOs just taking data indiscriminately from the Internet just because it’s indexable by Google.
I don’t think something is stolen if it’s analyzed and used for something new. It never matters if you came up with an idea, only what you do with that idea.
I think there is so much ignorance around this topic that people are so freely throwing around like it’s gospel. It’s frightening to see our society behaving this way about some relatively mundane (in the scheme of all artificial intelligence) technology and how quickly people want to kill something new without understanding it’s fundamentals.
https://www.eff.org/deeplinks/2023/04/how-we-think-about-copyright-and-ai-art-0
So anyone who creates something remotely similar to something online is plagiarizing, got it.
Folks, that’s how we all do things - we read stuff, we observe conversations, we look at art, we listen to music, and what we create is a synthesis of our experiences.
Yes, it is possible for AI to plagiarize, but that needs to be evaluated on a case by case basis, just as it is for humans.
