| lmao the problem is not that chatgpt is storing PII of the people chatting with it (ie you and me who are using the chatgpt website). this is not that the italian authorities complained about. the problems are 1. that the entire gpt3.5 model (and probably all models of openai) have a vast amount of PII stored from the scraped training texts, and 2. chatgpt gets a lot of the info wrong/hallucinated (ask it about yourself and see what is spews out). both are currently unfixable problems (without a breakthrough in science, not impossible but hard). you cannot surgically modify the knowledge of an LLM. it's probably only realistically fixable by lobotomizing chatgpt into langchain based search or calculation assistant that retrieves information from a mutable store, much like bing retrievs info from the bing index (which can be surgically modified upon receiving a GDPR request from a EU user). but chatGPT creating poems etc is fundamentally incompatible with EU bureaucracy and i am very curious to see who will win! |