Comment by fallingknife | Hacker News Reader

Comment by fallingknife | Hacker News Reader

> Here, Microsoft and OpenAI haven’t just used The Times’ content to teach the AI how to communicate, they’ve launched news-specific services and features that ingest both archived content and brand new articles from The Times.

What is this referring to? AFAIK reporting on real time news is beyond the capability of an LLM, and a Google search doesn't turn up any news reporting feature.

> But the hallucinations aren’t facts; they’re lies. And even if the defendants prevail in arguing that the AIs are mostly just providing people with unprotectable facts, there’s very little to shield them from liability for the lies, both with respect to trademark dilution claims, but also with respect to potential libel or privacy-related claims that might be brought by other individuals.

This seems really weak because libel requires damages. It would be incredibly hard for the NYT to prove any damages from ChatGPT lying about it. Also libel requires "actual malice" which means the libeler needs to know that their statements are false. Good luck showing that with an LLM.

> libel requires "actual malice" which means the libeler needs to know that their statements are false

Creators of LLMs (OpenAI in this case) should know that their LLMs regularly provide false statements.

One could argue that the LLM are not making any statements at all, as that would require intent and purpose beyond what they can (currently) do. The LLM outputs the most likely token to appear next, based on what it learned during training. This could be anything, and it happens to generate some text that sometimes looks like statements, but is just text generated by lots of math.

Of course I'm not actually arguing this, and have no horses in this race. But not hard to imagine that someone could make this argument and also believe it.

Telling a lie is not illegal by itself. Telling a lie with the intent of harming someone, especially their reputation, is another story.

And the NYT knows that its articles frequently contain false statements too. In this case they were sued for it and won even though they admitted in court that the libelous statements were not true. https://www.nytimes.com/2022/02/15/business/media/new-york-t.... The actual malice standard is next to impossible to meet.

> the libeler needs to know that their statements are false

That's only the first half of the or conditional. The second half is "with reckless disregard of whether it was false or not." That's very different, and I'd think the model and it's creators not caring if the output is true but marketing it as usually true is having a reckless disregard to the truth of it's outputs. There's no checking if it's real.

LLM hallucinations are a well known problem, and accepted as such by OpenAI etc... perhaps, as a group / statistically, that rises to the level of "knowing statements [can be] false"? of course, IANAL... but keen to see how it plays out.

>Also libel requires "actual malice" which means the libeler needs to know that their statements are false. Good luck showing that with an LLM.

Isn't that easily proven by the text at the bottom of chatgpt: "ChatGPT can make mistakes. Consider checking important information."

Disclosing that the information is prone to error is the opposite of a malicious lie

ChatGPT contains the disclaimer, proving that OpenAI are well aware of the tendency of the model to confidently misstate fact. Bing chat contains no such disclaimer, nor does the OpenAI playground that many use as a product itself.

I’m not really interested in following every service but… idk. Is that actually true? I doubt it.

I just read the complaint and the question we are trying to answer here is moot. This part of the complaint hinges on verbatim copies of their articles being intermingled with fake excerpts from fake articles making it impossible for a reader to differentiate the two.

If I wrote 10 fake headlines, I might be protected. If I copied 9 from NYT and then made one up, it's a lot harder for me to argue I didn't intend to deceive.

Bing/Copilot, as well as Web Browsing in ChatGPT, can look up the news with simple web searches, and (allegedly, according to the lawsuit) bypass certain paywalls doing so. Even when not bypassing paywalls, there's still issues with hallucinations (sometimes I click on a link and it's not even remotely what Bing claims it is) and stripping out things like attributions or referral links (e.g. NYT's Wirecutter makes money through affiliate links).

That they're not part of the main GPT-4 model itself is true, but the lawsuit is against OpenAI and Microsoft, not vs. GPT-4.

Reply on news.ycombinator.com