When people edit a chat transcript to make it 'safe' to share, does their editing reveal what they really think is private?

This explores whether watching people make privacy calls in the moment, while they edit, shows what they actually register as sensitive, compared with asking them afterward how safe they feel. The corpus has no study built around capturing those in-the-moment judgments. It does have strong evidence for why such a study is needed: what people say about their privacy and what they actually remove are often two different things.

The clearest case comes from a 188-person study where participants cleaned up chatbot interview transcripts before sharing them Why do people feel safe sharing when they haven't actually removed sensitive information?. People who edited by hand felt just as confident about their privacy as people who got AI suggestions, but they left far more identifying details in their transcripts. Their comfort came from the editing interface, not from how much they had actually exposed. So a post-task survey would have measured the wrong thing. The edits themselves, meaning what got deleted and what got left in, are the real record of what people noticed. Watching those edits as they happen is the only way to separate "I saw this and kept it" from "I never saw it."

That second case, never seeing it, keeps coming up across the corpus. In a 300-person survey, people chose AI-written replies that leaked private information over their own drafts 55% of the time Do people notice when AI agents leak private information?. The researchers put this down to people not spotting the leak, not to a deliberate trade-off. In a study of LLM memory, more than 30 of 40 users never noticed what the system had stored or inferred about them. Once it was shown to them, every participant wanted controls to edit and delete it Do users notice when LLMs infer and retain private information?. Together these suggest that people's sense of what's sensitive depends heavily on what's put in front of them. Ask them in the abstract and they underreport. Show them the actual item and their concern switches on.

There's also a clue about where attention goes. An analysis of Reddit posts about ChatGPT found that users worry mostly about what data gets collected (43.7%) and much less about how it's used or how long it's kept What privacy concerns matter most to ChatGPT users?. If that pattern carries over to editing, people would probably cut obvious identifiers like names and addresses and miss details that only become identifying in context or when combined. That's exactly the gap real-time observation could map. Another strand points to a design fix: in MemoAnalyzer, users state their privacy preferences and the AI carries out the edits Does splitting privacy work between users and AI reduce cognitive burden?. This takes the job of spotting every sensitive detail off the person, though the study hasn't yet shown that it improves what actually gets protected.

The surprising part is that the act of editing may itself create false confidence. Research on AI-written text finds that having more control over the text increases people's sense of ownership of it Does user control over AI text shape feelings of ownership?, and the transcript study shows the same effect for privacy: hands-on editing makes people feel protected whether or not they are. So real-time privacy judgment is valuable less as a window into what people notice and more as a way to measure the gap between feeling careful and being careful. Expect that gap to be large, and to be widest where people feel most in control.