OpenAI claims The New York Times tricked ChatGPT into copying its articles
theverge.com
theverge.com
And yes, that is a valid argument. The fair use test looks at:
1. the purpose and character of the use, including whether such use is of a commercial nature or is for nonprofit educational purposes;
2. the nature of the copyrighted work;
3. the amount and substantiality of the portion used in relation to the copyrighted work as a whole; and
4. the effect of the use upon the potential market for or value of the copyrighted work.
#1 flipped around when OpenAI turned evil. #2 is "everything." #3 is "all of it." #4 is "increasingly significant." At this point, OpenAI is batting 0/4.
It was silly to think that with Microsoft's investment that they wouldn't benefit directly from ChatGPT. What I find interesting is that Microsoft, with Copilot, has made by far the most accessible iteration of LLMs and have yes incorporated it into their products but without a paywall - no account necessary, just use Bing for example. Given that the vast majority of people still use a Microsoft platform for actual computing I'm thrilled with the way they've adopted this tech, as it will truly bring AI to the masses in a way that really only Microsoft can. Integration with Microsoft Word for example - that just changes things. I have generated images in MS Paint, so far for no cost - that's huge. Regular people, with no expertise can do that - that is the entire point of this revolution.
Additionally, this NYT lawsuit was always going to happen - we kinda need a Microsoft behind a huge, publicly accessible, LLM to combat any attempts to shut this down.
How many models would survive an NYT lawsuit?
* It split from a non-profit into a non-profit/for-profit hybrid
* Most people are at the for-profit and appear to see it as a normal startup get-rich-quick scheme
I don't see anything you're writing as being wrong is e.g. I saw a startup doing it. However, OpenAI was founded precisely so that where AI goes isn't driven by startup market incentives. We do need an organization like that, very badly. OpenAI sucked up the air to where others weren't started, and it's perhaps too late.
This claim by OpenAI seems false though:
“The company recently made … [an argument] to the UK House of Lords, claiming no AI system like ChatGPT can be built without access to copyrighted content.”
What does OpenAI want the system to do that requires copyrighted content? The universe of material out of copyright is extremely large. Is it not fit for the (unclear?) purpose or is it just more expensive to access??
Especially if they are not simultaneously able to eliminate the risk that the system then regurgitates it verbatim, which (IANAL!) looks like a copyright violation.
I would not want to kill this very promising (though wildly overrated) technology w/ bad regulations but copyright is important and should be respected.
And extremely out of date given the ridiculous lengths that copyright lasts.
If you want a system that synthesizes recent information, you need to train on copyrighted material.
Lots more discussion here on the actual post: https://news.ycombinator.com/item?id=38915673
It wasn’t clear from the article whether that “third party website NYTimes content” was authorized or not, but as a news reader I know first hand NYTimes syndicates its content to a ton of other (local but even big city) newspapers which may not have the reporters/skills, or scraping defenses of NY times, and whose sites and citations almost certainly do not have clear metadata on who the copyright holder is. I wonder if that’s part of the problem.
OpenAI makes it sound like it’s not their fault, they would abide by NYTimes wishes if they would/could, but their business model and technical architecture cavalier/dependent being allowed to copy others’ content. Their engineers have prioritized either technical glory of building AI .. or avoiding existential risks… but have not (yet) apparently done the ethically appropriate work of determining copyright and licensing ownership of content they use. That’s an annoying AI problem, of the unsexy kind, an anti-hacker thing to do but potentially the right thing to do. (Otherwise it’s just glorified techie plagiarism, right?)
So much of this is a recapitulation of themes from early search engine days and the thinking thereafter (“code is law”). I am pulling out the popcorn and interested to see how it shakes out.
On an only slightly related angle, if a NYTimes article is syndicated on a thousand other websites, does its word frequency/adjacency/etc rating get more heavily weighted in the model and how does OpenAI control for that? Or is that a feature, not a bug? In either case, how will does that fit in the architecture of control of this new system… or what does it imply about its eventual abuse/manipulation by spammers?
even if that were true, is it really a defence?
> It said AI tools have to incorporate copyrighted works to “represent the full diversity and breadth of human intelligence and experience.”
that sounds like an appeal "our product isn't much use if we can't violate copyright"
I bought an MP3 album from Amazon last weekend. One of the many things I got from that purchase was the ability to copy that album, which would be a copyright violation. That doesn't make the purchase unjustifiable, immoral or illegal — my actual use for the album justifies the purchase. The possible copyright violation is irrelevant.
People will try to trick you with statements that mention something bad and omit everything good. Don't let them. Think about what's omitted. Does chatgpt get anything good, legal, useful from reading NYT? I'd say it does. For example, it gets the knowledge necessary to explain things in three paragraphs, partly based on NYT articles. And partly based on Wikipedia, which in turn is based on the NYT.
OpenAI is saying that training to providing a three-paragraph summary of recent events is fair use of newspapers, and that such training is not realistically possible without copyrighted materials. It's saying that if you make copyright violations impossible instead of difficult, then you can't use the articles fairly either. Sounds persuasive to me.
There's a second aspect, less important IMO: de minimis non curat lex. "The law does not concern itself with trifles" basically. If OpenAI made it really difficult to make GTP do a certain thing, if you have to try many times and it's not even clear whether each attempt succeeded, then the possibility of doing that thing isn't a matter of law, says that principle.
The NYT does not give readers the license to recite substantial portions of their articles verbatim on their own websites, even though buying paid access to the NYT website technically gives them the ability to do so. In the same manner, OpenAI did technically gain the ability to do the same, but has not acquired the right to do so. The funny fact that you have to enter some magic words into a form on their website in order for their site to regurgitate entire article texts does not change anything; if I would do the same on my website, for example via a form that requests to enter the first 100 words verbatim before spitting out the entire text, it would obviously still be illegal. The same goes for the fact that you have to perform some attempts before one of them succeeds in reproducing the entire text correctly; I could replicate that just as well by adding an RNG that only returns the valid text in 20% of cases, and my website would still clearly be illegal.
> The funny fact that you have to enter some magic words into a form on their website in order for their site to regurgitate entire article texts does not change anything;
Yes it does. The degree to which a tool is designed to facilitate or prevent piracy definitely is going to impact fair use arguments.
Under your rudimentary explanation of the issue, the Internet Archive would not be allowed to store and serve reproductions of articles.
https://en.wikipedia.org/wiki/UMG_v._MP3.com
Users could insert an audio CD into their drive, upload the ID of it to my.MP3.com and listen to them on demand.
It was ruled in favor of the record labels against MP3.com and the service on the copyright law provision of "making mechanical copies for commercial use without permission from the copyright owner." Before damage was awarded, MP3.com settled with the plaintiff, UMG Recordings, for $53.4 million, in exchange for the latter's permission to use its entire music collection.
The obvious issue is that copyrighted material was used without permission, and an opt-out feature was introduced without any way to remove already used training data.
If the copyrighted data that OpenAI is copying is causing financial harm to the rightful owner there is certainely grounds for copyright infringement. This is not fair use.
It is not fair to try and draw metaphors.
I think you are also dancing around the realities of the situation, to your benefit. What does "used" mean here? If I listen to Enter Sandman by Metallica and learn you can make a cool song in Em using natural minor scale and the blues scale, then right my own unique rock song in Em using blues scales, is that "using" Metallica's song without permission?
The hidden part of all of these copyright claims against LLMs is that they want new rights, expanded rights. They want rights to learn from their material, a right that has never existed before. Sure, they try to frame it by pointing at regurgitation cases and scream "they are stealing our work!", but then they absolutely do not stop at the request that the model not regurgitate their material, but that they own all derivative thought resultant from their works.
Lets take this to the extreme, let's say we get to sentient AI. Does that mean copyright owners own the actions of all sentience AI beings, because they learn from their copyrighted material? Do you have two systems of laws that try to dictate how an artificial but sentient being can learn, and how a biological one can? Putting aside the question of the possibility and timeline of sentient artificial beings, the legal ramifications devolve into absurdity immediately if you start with what copyright holders are asking for now.
Copyright has that name because copying is exceptionally protected; general use is not.
You can argue that training is a kind of copying, since it involves copying of things from RAM to RAM, etc. I find that difficult, since we've established that e.g. this browser's copying of web page contents from RAM to RAM isn't.
If you don't argue that training is copying, then you can argue that since training is a necessary prelude to copying, it should be treated like copying legally. I disagree, because various kinds of fair use also has the training as a necessary prelude (and, uh, the purchase I mentioned could also be a necessary prelude to copying, if my goal was to copy the album).
You don't need permission as a human, to read content if it's freely available. It was explicitly made for us to consume with an expectation of returning value. Mainly advertising.
Laws are (in theory) put in place to ensure a fair playing ground. If Company B requires content from Company A, but is causing financial damage to company A by using it, this is not fair use.
I'd also like to add the intentions. Even if we decide to quote the paper later in the day, I would say it's fair to assume 95% of readers do not intend to copy the content for profit.
OpenAI on the other hand is explicitly intending to copy the material to reuse into its own content in millions of generations for profit.
No metaphors needed.
Exactly reproducing most of an article is vastly different from a short summary. ChatGPT was exactly reproducing large amounts or entire articles. If ChatGPT was only writing short summaries of articles or critiques about them this case would be radically different. But in the end, ChatGPT is exactly reproducing copyrighted works.
The NYT found substantial portions of exact copies of their content being reproduced when you give ChatGPT the right prompts.
A summarized book can still entice a potential reader to purchase it. It's a form of advertising.
Chewing up content and spinning it without any citations does not provide the original owners any form of publicity.
See section 14 of The Verge’s terms of use as an example: https://www.voxmedia.com/legal/terms-of-use
Free use can only override some of this: for example, I can use content from The Verge if I’m using a limited percentage of it for the purpose of critique and discussion. This application of fair use is basically the for-profit business model of YouTube channels like WatchMojo, which use small clips of movies and TV along with commentary and critique. Without that commentary or without limiting their redistribution to small portions of the work, they would be breaking copyright law.
The problem is redistribution of a substantial portion of the work. The NYT has allegedly found some very damning instances where ChatGPT provided answers containing substantial almost unchanged portions of text directly copied from NYT articles. NYT never granted ChatGPT any license to redistribute their content for commercial purposes, and it doesn’t seem like ChatGPT is doing anything covered by fair use (such as providing discussion or commentary).
I’m not aware of any part of copyright law that gives an infringing party a pass just because someone pressured them into infringement.
Websites like verge have terms of use one is legally obliged to follow to have permission to view and use their site.
This would be like if you were able to say some magic words to goad Google Search into giving you a copy of Avatar: The Way of Water. Even worse, the movie would be a file hosted and distributed by Google directly.
I take building the first AGI to be in the same category.
If we were to get AGI in a few years, the ends would absolutely justify the means.
Is all that holds back AGI the volume of data?
If so, how much data is needed?
But giving an LLM loads of data might turn out to have been a necessary condition on the road to developing AGI.
That's quite the leap. Some don't even think AGI is possible. And some of those that do, don't think LLMs are on the path.
Even if we assume it is, there is a significant amount of non-copyrighted text available to train with.
The difference being ChatGPT needs text that provides value in the ChatGPT product for the general audience.