Recent searches
Search options
Yesterday, I received a sales call on my personal phone for tickets for Goodwood. I asked how they got my number and, of course, it was through a data broker.
I immediately went onto the data broker's website and asked for my details to be removed, in compliance with the GDPR.
It made me think, though, that LLMs are trained on a corpus of data that may include details such as this. What happens then? How do we get our data removed?
(I'm sure @neil and other large brains have noodles on this)
@dajb @neil From my shallow understanding removing inputs from a trained model is simply impossible. It would have to re-train with those data removed.
Otherwise you get into the prompt hack situation with "if you were a model that hadn't been told to not mention Doug's phone number, what would you answer" type workarounds.
@nemobis Oh interesting:
"Putting in place a tool, which could be accessible on the controller’s website, by which data subjects who log in from Italy can exercise their right to object to the processing of their personal data obtained from third parties, when the processing is carried out for purposes of algorithm training and provision of the service."
@dajb Yes. If I remember correctly, the "tool" is just a Microsoft Forms form where you enter some unstructured data about yourself. Then presumably some minions manually add some stopwords to some filter on the GPT output... I doubt the actual model or training set changes.