[jug]

Hmm, it's not that simple, is it? Let's say the AI is trained on the tweet "Ben Adams drove to Mexico yesterday but I still haven't heard from him."

From this knowledge, you can ask the AI "Who has driven to Mexico" and it might know that Ben Adams did, and reply with that.

HOWEVER it's also baked into the model and can't be surgically removed after a complaint. That's the irreversibility part. You can't undo isolated training. You need to provide it a new data set and train it all over again. They won't do that because it's too costly.

The problem with the above example is of course that it can also contain sensitive or private user details.

I've easily extracted the complete song lyrics to the letter from GPT-4 even if OpenAI try to put up guardrails against it due to the copyright issues. AI is really still in the wild west phase...