404 Media reported on Monday that hundreds of contractors read real ChatGPT conversations under a project called Lily. Reviewers see the whole chat plus a memory summary of what you asked before, then score replies one to seven. OpenAI says a filter strips personal details first, but admits it can miss them.