Browsing: Data Privacy
A Reddit user uploaded a screenshot of an Excel sheet and ChatGPT described a shirtless man in photographic detail, then defended the claim when challenged. Here is why vision models do this, and how to tell a hallucination from an actual data leak.
Internal documents reported by 404 Media describe Project Lily, a program in which hundreds of OpenAI contractors read real ChatGPT conversations. Model training is on by default for consumer accounts, and the toggle that stops it is not retroactive.