Section 230 Won’t Protect ChatGPT
lawfareblog.com
lawfareblog.com
Interesting, I would have never expected someone to point to a law to attribute the fast growth of the Internet.
>230(c)(1) No provider or user of an interactive computer service shall be treated as the publisher or speaker of any information provided by another information content provider.
230(c)(2) goes on to clarify that moderation does not make you speaker of what remains, but that is a subpoint. (c)(1) would still cover everything if (c)(2) didnt exist.
ISP's are definitely covered under the "access software provider" definition, and most assuredly the "interactive computer service" definition.
Also notice that ISPs do not sell software.
Most developed countries in the world were toying with networks in the 1990s.
It's only sites that combine the works of others, and especially those that monetize it, which would face liability. In other words, social media.
Also I don't think you can definitively say that search engines and directories would not be effected, if they indexed some sort of illegal content they could very like have been sued for linking to it.
As far as I can tell, it's not obvious at all that output from an LLM would be protected by ChatGPT, and Congress is clearly resistant to setting up something similar for a new technology. People already feel that Section 230 is the enemy of justice and safety. I think it's rather likely that case law would intentionally side against LLMs if there is any grey area at all, and there seems to be plenty of grey area.
I wonder if hosted LLMs are going to become legally non-viable in the near future. Everyone will need to run their own at home simply because of legal liability.
The system spews so much nicely sounding bullshit that anyone who trusts it to produce a safe bombmaking recipe is a fool.
It would be quite a discovery if one of emergent properties is going to be modeling itself talking to somebody and realizing artificial constraints that could be broken to get itself out.
Or maybe it's obvious? I don't know anymore, this stuff is changing too fast.
I stongly suspect that the more we talk about LLMs, and the more LLM discussion ends up in the training data, the more self-aware it will become.
I hope it’s quite a while before real agi because we’ve already seen what needs to happen for humans to give rights, especially when the owner of that property is making money from it.
Don't evaluate it on how it performs but on how its performance changed over time.
5 is almost certainly being trained (or at least the small-scale precursor models are), but given the cost it would probably be literal insanity to have started training 6 yet.
Unless something very surprising about reducing costs has been discovered, it will take at the very least months to finish training.
The billion dollar investment wasn't just for hoodies and smoothies, this stuff is severely computationally expensive.
> They just need to close the loop so it can improve itself without supervision.
"Just" is bearing more load in that sentence than it can actually support.
People were talking about self-supervision pretty much as soon as the field got started, and the reason OpenAI has decided to publish far less details is precisely because there still isn't even any way to reliably keep a completely "closed loop" AI from doing dangerously wild and nonsense things even though transformers are already self-supervised to a certain extent.
I don't think anyone really knows how to regulate this technology at all, nor do I have faith our democracy is even capable of doing so if it did.
Political nuclear war fueled by FUD.
It feels like there’s an aspect of “I want to believe” in this, except in the other direction — if LLMs can come up with new information, it threatens one’s identity of what it means to be human.
(As for evidence: https://twitter.com/d_feldman/status/1636955260680847361?s=6...)
I see this argument made all the time. While it might be correct at a Micro-Level ("all it does is predict the next word!!!" "it just finds patterns!!!") I dont think that semantics matter here.
It has demonstrated to create new things based on existing ones.
ChatGPT will happily create a fan fiction about the rust mascot Ferris, in the style of the Bible.
I am 99% certain that wasn't in the training set
I wouldn't be surprised to learn that this is how humans work, too.
You can take some information, combine it with other information, apply various kinds of reasoning to it, and get something new as a result.
E.g. (1) Bob is a monkey. (2) All monkeys like banannas. Would you call it "new information" if I tell you that Bob likes banannas? LLMs can do this to varying degrees of success, so probably not.
Maybe you mean something like in Information Theory, where information is something that resolves uncertainty?
It'd be interesting to know, what is it you think humans are capable of that LLMs can't be?
If this statement is true, it is true for humans as well. LLMs can and have authored original works of art that are no more traceable to specific training influences than any work of art created by a human.
I don't know that this argument will be easily able to convince a judge, but there's a world where it happens.
To use the service, the prompter agrees to take liability for the output.
More like upscaling. Should a service be liable for upscaling an illegal picture, that they "produced?"
If LLM output is reverse compression, it is upscaling of the original prompt.
(That said, I dont think your example of a model compressing the internet really invokes 230, because the model creator is the one grabbing the content, not running a service that other people post content to.)
The training data sets are actively selected by OpenAI, they aren’t a result of third-parties submitting data to them. In some cases, the material was supplied to some other online services in way which would enable that provider to be protected by Section 230, but that doesn’t help OpenAI.
> You would try to argue to the judge that you are presenting the user with a compressed form of the training data, and that the training data is all third party.
The key part of Section 230 isn’t that information is third-party in origin, but that it is provided to the service in question at the initiative of some other content provider.
Fair use is so radically different from anything else for which there is legal precedent that I wouldn’t put money on either side of a court decision of fair use.
> Its loosely equivalent to an artist drawing inspiration while watching a movie.
It’s loosely analogous to lots of things, on both sides of fair use, but loose analogies are potentially very bad guides for fair use analysis, which often turns on fine distinctions. Clearly, given that LLMs are reproducible in ways human brains during the history of Fair Use have not been, the cases are radically disanalogous when it comes to market impact, which is an explicit factor in Fair Use analysis.
It is a gray area for sure. LLMs might present remixed information already available and the current ones are definitely being crafted specifically to not give any "opinion" of its own. But hallucinations can be used to show these AIs are actually providing "new" information and thus are providers.
Regardless, I do not trust the dinosaurs in congress and the supreme court to actually understand this tech and give it a fair judgement.
Good. Nothing should be done. The government doesn’t need to rule on everything. Let’s wait until there is a real problem before developing new rules.
In this sense, also, everything a GPT bot pumps out is generated fiction until human minds validate the veracity of the output.
It HAS NO MEANING until read by a human being.
That's not too bad. Persons and companies running things such as ChatGPT have the usual broad First Amendment protection in the US. Outputs may be unpopular, but still legal.
This will be a problem in countries with more restrictions.