> While unlikely, we cannot rule out that de-identified data derived from their usage of our products helped improve our models
That basically means, we don’t know, and we hope the model didn’t look up user conversations, and the best thing we can do is hope.
That’s seriously disgusting. I can understand why on a technical level why perhaps it is impossible to answer what exactly the model had access to, but it still is disgusting.
it just means that they've been working with paraphrased user data everywhere, which any smart person can figure out is how anthropic and openai train on so called non-retained data.
This is more of an expectation of privacy issue. If you write on an envelope and USPS has a copy, you have no reason to be mad; if you write on a letter inside the envelope and USPS still has a copy, you could rightfully be mad and say this is disgusting.
If the work is duplicative/derivative then the preprints they put in sessions can be shown by the users and we can see.
Nobody would be disgusted at following stated policy, that's doesn't make sense, without first objecting to the policy. But you're bringing up distractions from the actual concerns: did OpenAI use the private chats and why won't they confirm or deny it?