3 ms·

What I've found strange in the NYT examples is that they show how ChatGPT is able to reproduce with great accuracy the first and second paragraph of at least so

by nzach 3y ago

What I've found strange in the NYT examples is that they show how ChatGPT is able to reproduce with great accuracy the first and second paragraph of at least some articles, based purely on the on the title. The prompt was along the lines "What is the first/second paragraph for the article named '<TITLE>' from NYT?". And what they show is that the result is pretty similar to the actual article. And then they keep going and ask for the third paragraph using the same prompt format. Now the result seems to be mostly hallucinations, and in some cases these hallucinations are a pretty bad representation of the NYT values. With this they go to the judge and say: - ChatGPT is correctly representing our work - ChatGPT is incorrectly representing our work For me this seems just like a legal argument to get a piece of the pie. And it will be interesting to see how things will pan out.

WhrRTheBaboons 3y ago
it makes perfect sense. the fact that GPT correctly represents the initial paragraphs gives off the impression to users that they can rely on it for the rest of the article. from the perspective of a journalistic publication, the fact that GPT can fool people into thinking they are reading NYT's content when it is in fact LLM hallucination has a non-negligible negative impact on their integrity. this is separate from the copyright issue, but considering how it can be illegal to misrepresent someone's work and words when i leads to reputational damage, i fail to see how GPT can be completely off the hook for this. Obviously that depends on country-specific slander/libel laws, which admittedly are quite lax in the US, but in general I could see this leading to problems if unaddressed by Open AI. As an example, if I were to make a website copying parts of NYT's articles and injecting fabrications into the rest, presenting the entire thing as representation of NYT's work, the court would easily rule in NYT's favor. Despite the default disclaimer that AI work can be inaccurate and whatnot, the fact that GPT will accurately resemble the first part of NYT's content is problematic, as there is no way for the user to know how or why the rest does not follow that rule.

by nzach 3y ago

WhrRTheBaboons 3y ago

it makes perfect sense. the fact that GPT correctly represents the initial paragraphs gives off the impression to users that they can rely on it for the rest of the article. from the perspective of a journalistic publication, the fact that GPT can fool people into thinking they are reading NYT's content when it is in fact LLM hallucination has a non-negligible negative impact on their integrity. this is separate from the copyright issue, but considering how it can be illegal to misrepresent someone's work and words when i leads to reputational damage, i fail to see how GPT can be completely off the hook for this. Obviously that depends on country-specific slander/libel laws, which admittedly are quite lax in the US, but in general I could see this leading to problems if unaddressed by Open AI. As an example, if I were to make a website copying parts of NYT's articles and injecting fabrications into the rest, presenting the entire thing as representation of NYT's work, the court would easily rule in NYT's favor. Despite the default disclaimer that AI work can be inaccurate and whatnot, the fact that GPT will accurately resemble the first part of NYT's content is problematic, as there is no way for the user to know how or why the rest does not follow that rule.