Hacker News Reader

Comment by satvikpendem | Hacker News Reader

If AI is ruled as fair use, no license will stop it, which it likely is to be, notwithstanding the NY Times lawsuit against OpenAI alleging that ChatGPT can reproduce articles wholesale.

View on HN

If it is not ruled fair use, or similar, all non-copyright countries are going to advance far faster. Whilst there are many ethical discussions to have, this will come down to the simple fact of there being people who will take the data anyway and thus have the advantage in datasets, assuming that matters.

I think the ideal of more data == faster advancements may turn out to be a false equivalency. Similar to prior innovations, a healthy headcount, headstart and steady R&D funding will more likely make a difference.

Besides, rather than vacuuming up exisiting data, future resources might be better spent on creating bespoke datasets with user consent and content licensing baked in from the start.

> Besides, rather than vacuuming up exisiting data future resources might be better spent on creating bespoke datasets with user consent and content licensing baked in from the start.

Not mutually exclusive.

Probably not, but I think it is worthwhile to see if models can achieve sota with less, and if not, have individuals clearly onboard and informed of their data contributions.

We have conducted large statistical surveys in the past, so I see little reason why these and other established data collection practices could not be updated and utilised for refining contemporary models.

For DNA and genetics more data is better. Data and it's interpretation is the bottleneck for some subjects but not all. Getting more data from search engines about purchasing habits might not yield more/better results as MoE showed when 2-3 experts was better than 4-6.

At first, data is needed to overcome scarcity in domains where it has traditionally been difficult to obtain -- but I don't think it's quite the same as the 'data' being (colloquially) referred to here. We may want to discern between user generated content and genetic features that one has little influence in creating, among the many other data types and kinds.

It's probably a large reason why this space is so difficult to license/regulate: data to one may simply be 'noise' to another, or not even register as such. Definitions unfortunately, do not always map.

The US shot itself in the hand foot and ass by letting the wright brothers patent troll so hard that they couldn't even make their own planes during WW1, they were buying American designs built by France.

https://www.centennialofflight.net/essay/Aerospace/WWi/Aero5...

> The United States did not produce any aircraft of its own design for use at the front during World War I.

If it happens again for AI who knows what will happen.

You're making an assumption there, that "advance" is just one axis. There are many ways a society could "advance", many of which are wildly impeded by current AI technologies. I'm just not sold that it's a good technology to let run free without any regulation. (Heck, not many are.)

There has already been one court decision that pushes against a blanket fair use protection for AI training: https://chatgptiseatingtheworld.com/2023/10/13/judge-bibas-i...

And I think there are dozens of unfinished AI copyright infringement lawsuits... with likely more to come, I speculate.

https://chatgptiseatingtheworld.com/2023/12/27/master-list-o...

https://heathermeeker.com/ai-copyright-et-al-litigation-scor...

I don't think that first link means what you think it means:

> Ultimately, Judge Bibas denied summary judgment and ruled the case should proceed to trial.

> Judge Bibas cautioned against “overreading” the Supreme Court’s recent fair use decision in Andy Warhol Foundation, in which the Court limited the weight to give a defendant’s transformative purpose if it is commercial and substantially similar in purpose as the plaintiff’s work.

> "Ross’s uses were undoubtedly commercial. And one of its goals was to compete with Westlaw. Thomson Reuters contends that this commercial use weighs heavily against finding fair use. [...]"

All that the judge is saying is that previous rulings are not comparable to AI related ones. It is also difficult to see how ChatGPT is competing with NYT, as ChatGPT does not generate (actual, real) news, because it cannot interact with the world as journalists can. Maybe image generators could but a similar judgment stated that it was difficult to see how AI generated images were similar to an existing artist's.

It's clearly competing with NYT if people get their news from NYT via ChatGPT instead of getting it from NYT directly.

Otherwise you could argue that if I copy and paste the NYT articles onto my own website, because I'm not doing any journalism, I'm not competing with the NYT - which is clearly ludicrous.

So then it needs to be proved that real people actually cancelled their NYT subscription in favor of reading articles from chatGPT. This is incredibly unlikely for several reasons least of which is GPTs training cutoff, the need to copy past several paragraphs of the article to get GPT to finish it among other reasons.

Fortunately, that’s not how copyright works. Burden of proof is on OpenAI to justify their copying without permission, they don’t get to copy anything and then demand that anyone who doesn’t like it prove that financial damage was done first before they stop. Financial damage does count under copyright, but is not required to win an infringement claim. Courts or the copyright office might decide that AI can count as fair use, or they might continue to allow suits every time an AI spits out verbatim copyrighted work in violation of the law, or we might get some entirely new criteria for copyrights. It’s going to be interesting.

To be honest I really dont understand how copyright works. I have read through the first several pages of the NYT case PDF but it still didnt make much sense to me. Is the issue just that GPT is able to repeat the article word for word or is the issue about the article being in the training data at all? If its not about losing business what is it about? Could you or someone else expand a bit on what exactly the problem is and maybe even speculate as to potential outcomes?

This is a good question, and I suspect the answer is some of both, but the lawsuit is very specifically claiming that ChatGPT is “memorizing” and reproducing NYT articles verbatim, and this is clearly in violation of the law. I suspect they’re also trying to push on the issue of it being illegal to train AI on the NYT archive, but that’s a bigger and more difficult question. The issue of verbatim reproduction makes a clear legal case, which is why they’re relying on that first, and letting that lead to the broader discussion. Exhibit J in the suit is titled: “One hundred examples of GPT-4 memorizing content from the New York Times”.

> It is also difficult to see how ChatGPT is competing with NYT

Why? NYT paid for and put work into their writing, and they do make some money on holding the archive of their work. Not everything they publish is breaking current events either. Chatgpt does not have to come up with new news in order to compete with NYT archive, their editors, their writing style, or their content, among other possibilities. AI’s very purpose is to absorb the content and style of its training data and be able to reproduce it more cheaply than the original. It doesn’t seem difficult at all to imagine many ways this does direct and indirect harm to NYT, or anyone who had to put effort or money into creating something valuable.

Style and similar sorts of content is not necessarily copyrightable. I could write in the style of some author a similar story to one of theirs, like a murder mystery in the style of Agatha Christie, but that doesn't make it copyright infringement.

Yeah that’s true, good point. It’s only infringement if you string enough of the exact same words together verbatim as the originals / training data. Same is true of content. The NNs aren’t yet good enough to avoid memorizing individual training samples, but something tells me they will be very soon.

> It is also difficult to see how ChatGPT is competing with NYT

If I'm interested in Dick Cheney shooting someone by accident, and previously I would have googled that and been sent to NYT reporting on the incident (and get shown some ads or sold a subscription) but instead I ask ChatGPT - that seems like a competing product to me.

And fair use is a multi-pronged test - ChatGPT's use of the data is clearly of a commercial nature, and the portion of the copyrighted work used for training is presumably 100%. So while Wikipedia can quote an NYT article (being both noncommercial and a short excerpt) the court could well consider LLM training more in the nature of music sampling, where at times it's been held that even minuscule two second excerpts, distorted and mixed in with lots of other music, still need to be licensed.

And what exactly would ChatGPT tell you? It'd make up some story about that person, but it's not real news. I don't see how that's competing with newspapers at all.

The fact something is not good does not mean it's not competition.

I may think TikTok is vastly inferior to the National Geographic TV, but they're still competing for eyeballs and ad sales.

Unless NYT proves the few people with a NYT subscription cancels NYT for a chatGPT subscription I don't see anything coming from it. Publicly scraping is what all search engines do. There also isn't any proof of intention, imagine if the US bans AI and china allows it because of the NYT.

Not just that, but in many jurisdictions it is already "fair use" (TDM exceptions). EU, Japan, South Korea, Malaysia, Singapore, etc. If the US bans it, it's only going to hurt itself.

Sources on this? As far as I know, no jurisdiction has yet ruled on AI fair use policy.

For Japan, I saw this: https://petapixel.com/2023/06/05/japan-declares-ai-training-...

It's not fair use per se which is why I put it in quotes, but more like training is basically exempted from copyright in law: https://www.reedsmith.com/en/perspectives/ai-in-entertainmen...

For commercial use, too. For EU though, commercial use needs to respect opt-out. Other countries can ignore opt-outs.

notwithstanding all the cases that could rule it not fair use? That's quite the hedge.

Yes, but it was due to their specific implementation where NYT showed that ChatGPT "memorized" the words to an article, shown by setting the temperate of the model to 0. Other AI implementaions might easily not have this limitation, so it's dumb to take one implementation and create broadspanning rules for the entire idea as a whole.

Reply on news.ycombinator.com