Researchers Extract Personal Emails from AI Model
-
Researchers were able to extract personal email addresses from GPT-3.5 Turbo by providing it with a small sample of verified contacts.
-
The technique was able to bypass some of the model's safeguards against providing private information.
-
OpenAI claims they don't actively seek out private data but won't reveal specifics on their training data.
-
Models like GPT can "recall" data they were trained on even if not designed to store it.
-
Experts warn there's no guarantee commercial AI models trained on internet data won't retain or reveal private info.