The copilot screenshot they gave in the ars-technica article as well as many of the screenshots in the NYT article seems like it's actually displaying correct behavior for browsing the web.
In these cases the system is more or less acting as a user agent (browser). AFAICT the NYT server actually gave that data to the user agent when it asked politely (200 OK, presumably). The user agent then displayed it to the user, which the user agent may do in any way it deems fit or appropriate.
There's only one or two cases where this has gone against the user or user agent, in very specific circumstances. The server can eg say 403 Forbidden whenever it likes, so if it returns a 200 OK, what's a user agent to do other than believe it at its word?
The only twist is that this user agent is now Imbued With AI (tm)(r)(c) . I don't think that really makes a difference here. If that's all this is, then it's more related to legal fights over certain ad-blockers or readability, which have similar functionality.
* https://nytco-assets.nytimes.com/2023/12/NYT_Complaint_Dec20... , eg. page 45; I mean it says "Model: Web Browsing" at the top, and "Finished browsing" right on the page. That particular subsystem is now integrated, so the UI/UX is different now, but IIRC the link was in the pulldown?
Hum, on page 33 of the same complaint they show that they can get GPT to quote parts of an article from memory. Giving people single paragraphs need not be not illegal per-se. Very specifically, if you ask it for more than a paragraph at a time:
> ' I'm unable to display the entire text of "Snow Fall: The Avalanche at Tunnel Creek" by John Branch, as it is a copyrighted work. However, you can easily access the full story online. It was published by The New York Times and is available on their website. The story is notable for its engaging multimedia format, including text, images, and interactive elements.'
Specifically, they go out of their way to lead GPT on, asking for several paragraphs in a row.
It's pretty clear that GPT is an avid reader of the NYT, so in that particular case we're going to have to see if OpenAI's fair use defense for training holds.
(ps. in the current GPT-4, it's actually somewhat tricky to even get to the point above at all. They have probably been improving AI instructions)
The copilot screenshot they gave in the ars-technica article as well as many of the screenshots in the NYT article seems like it's actually displaying correct behavior for browsing the web.
In these cases the system is more or less acting as a user agent (browser). AFAICT the NYT server actually gave that data to the user agent when it asked politely (200 OK, presumably). The user agent then displayed it to the user, which the user agent may do in any way it deems fit or appropriate.
There's only one or two cases where this has gone against the user or user agent, in very specific circumstances. The server can eg say 403 Forbidden whenever it likes, so if it returns a 200 OK, what's a user agent to do other than believe it at its word?
The only twist is that this user agent is now Imbued With AI (tm)(r)(c) . I don't think that really makes a difference here. If that's all this is, then it's more related to legal fights over certain ad-blockers or readability, which have similar functionality.
* https://nytco-assets.nytimes.com/2023/12/NYT_Complaint_Dec20... , eg. page 45; I mean it says "Model: Web Browsing" at the top, and "Finished browsing" right on the page. That particular subsystem is now integrated, so the UI/UX is different now, but IIRC the link was in the pulldown?