How? This is retention for legal risk, not for training purposes.
They can still have legal contracts with other companies, that stipulate that they don't train on any of their data.
They can still have legal contracts with other companies, that stipulate that they don't train on any of their data.
There's a qualitative difference resulting from quantitatively much easier access (querying some database vs. having to physically look through court records) and processing capabilities (an army of lawyers reading millions of pages vs. anyone, via an LLM) that doesn't seem to be accounted for.
Of course I haven’t sent anything actually sensitive to ChatGTP, but the use of copyright law in order to enforce a stricter surveillance regime is giving very strong “Right to Read” vibes.
> each book had a copyright monitor that reported when and where it was read, and by whom, to Central Licensing. (They used this information to catch reading pirates, but also to sell personal interest profiles to retailers.)
> It didn’t matter whether you did anything harmful—the offense was making it hard for the administrators to check on you. They assumed this meant you were doing something else forbidden, and they did not need to know what it was.
It's one thing if you get pwned because a hacker broke into your servers. It is another thing if you get pwned because a hacker broken into somebody else's servers.
At this point, do we believe OpenAI has a strong security infrastructure? Given the court order, it doesn't seem possible for them to have sufficient security for practical purposes. Your data might be encrypted at rest, but who has the keys? When you're buying secure instances, you don't want the provider to have your keys...
Look at it this way. If you your phone was stolen would you want it to self destruct or keep everything? (Assume you can decide to self destruct it) clearly the latter is safer. Maybe the data has been pulled off and you're already pwned. But by deleting, if they didn't get the data they now won't be able to.
You just don't want to give adversaries infinite time to pwn you
Do search engine companies have this requirement as well? I remember back in the old days deanonymizing “anonymous” query logs was interesting. I can’t imagine there’s any secrecy left today.