NYT vs. GPT: Microsoft finds itself on the other side of an industry | Hacker News Reader

NYT vs. GPT: Microsoft finds itself on the other side of an industry

geekwire.com

44 pointsjatwork17 comments

Don't news publishers have an even harder time making a case here than (say) software publishers?

At least with software, there are explicit licenses governing how you're presumably allowed to use them. Is the New York Times going to claim people are agreeing to a license before reading their news?

>Is the New York Times going to claim people are agreeing to a license before reading their news?

They don't need to. They have implicit, automatic copyright on the published original works. The lawsuit shows examples of GPT using the NYT work verbatim.

To discover than a particular LLM is able to output verbatim a sections of NYT works when queried in certain ways is different from explicitly publishing a set of NYC works... i think.

OpenAI does not distribute it models, so the violation of copywrite if found to be a violation will be limited to the amount of people who queried the LLM in the same way that NYT did.

My comment was directed solely at the part I quoted regarding licenses -- the NYT does not need to claim anything about licenses because they have implicit copyright.

How that copyright shakes out in the end (what your statements focus on) is yet to be determined.

You completely missed my point.

I'm saying that between "you're violating copyright law" and "you're violating copyright law and you supposedly agreed you a license that contractually prevented you from doing this", the second sounds like a stronger claim.

The reason the current license ecosystem exists for open-source software is that there's a huge difference between owning the code and owning the binary distribution. This difference doesn't exist for newspaper articles (or books, visual art, etc). Even technical manuals are not machines in the way that software is. The fact that software is both text and tool means the legal issues are different and need clarification.

So existing copyright law was sufficient for written media, but didn't quite work for software. OTOH it seems that GPL-style copyleft licenses are being used more by writers and artists, so it's not like these issues exist in a vacuum. I expect copyleft will be more appealing to artists going forward given Midjourney et al.

I understand their claim to copyright of recent material but the real reason so much of their information was used is bc we are talking about the, essentially historical, NYT archives that date back to likely the early era of the newspaper itself.

I don't think they should have a right to the records of people who lived before simply bc they underpaid, over worked human journalists to exploit their creativity - largely to make the "Newspaper Barons" rich.

I don't think they ought to have implicit, automatic copyright on older stuff. How old is my only debate.

What good is old news really? Real old is good for college papers but news from 4 years ago... who cares?

This battle will be fascinating.

No it wont lead to anywhere. They will just get rid of the NYT bits of the corpus, then settle or just flat out not care if they lose. Transformers are relatively easy to train they arent GANs, the bottleneck is hardware and OpenAI has those bits sorted out.

Presumably, the outcome will start the series of legal precedents regarding copyright and GPT (and other AI stuff) training potentially going far beyond OpenAI and NYT.

In the lawsuit, NYT pointed out that their articles are the third-most used source in GPT-3, behind Wikipedia and the US patent office. We don't know what they used for GPT-4 but I imagine NYT is still very prominent. High-quality newspaper articles are valuable for training LLMs because they contain good prose and an enormous variety of factual information. They can't be replaced by throwing in a bunch of Reddit comments and hoping for the best.

"NYT bits of the corpus" is understating things - I imagine getting rid of NYT articles would significantly impact the overall quality of GPT-4's responses in many unexpected domains.

I very much doubt it'd be a problem. OpenAI's valuation is such a large multiple of NYT that they can afford to buy out a dozen or so publishers of NYT's valuation to get content to replace (or "just" license content from far more) if NYT itself isn't willing to play ball cheaply enough.

The NYT might be better than average, but it's not so special that you can't replace it.

The long-term effect if NYT wins this case will be that the "established" players and companies with deep pockets will get a massive moat protecting them against open models and new entrants - in the long run it might well be well worth whatever a loss will cost them.

The US is placing a big bet on AI to break the economic secular stagnation of recent decades. If IP is an obstacle to that, it would be great if IP laws are loosened.

This is what I expect/hope will happen.

The NYT could prolly win under current law - due to the interpretation of how GPT is reproducing the articles. The judge that may ultimately decide the future of how AI will fall for now in the US, will likely know very little about AI/LLMs.

I don't think reproducing a current article is that big deal tbh - it just retrieved the info from itself rather than the NYT bc it had already saved the info before and can't forget things. That's humanely possible by anyone with a photographic memory - it's more similar to recollection than infringement. The NYT is pissed the LLMs learned from their data - you can't stop something from learning - especially after the fact.

Obviously, the user intent not machine intent is key also. No infringement can happen without a prompt that suggests copyrighted material - easy to do or not, requires a person nonetheless.

The technical aspects of this is where they have a chance, taking a step back, it all just looks kind of silly.

Seriously, just consider this: I memorize a textbook, verbatim, about WWII and you have a term paper on WWII so you ask me to write some notes about the highlights, so I write notes that I take write from the verbatim text in my head and give them to you.

You use the notes I gave you verbatim in your paper and are found guilty of plagiarism.

You obviously didn't realize the notes I gave you on the spot were from a textbook - that wasn't what you were looking for. I didn't realize you were just going to use whatever I wrote in your paper.

Did I commit plagiarism or did you?

If the NYT doesn't settle this could set precedent that will have massive repercussions on how all AI companies collect data. This could end up being a hugely important case.

Even if they do, settlements often establish de facto precedents that corporate lawyers (and insurance companies) will strongly advise you steer away from.

Reply on news.ycombinator.com