If people can freely edit their own chat transcripts, why do they leave more personal details in than when AI suggests the edits?
Does default inaction explain why free-editing participants left more personal information behind?
This explores whether people who were free to edit their own chatbot transcripts left more identifying details in simply because doing nothing was the easy default, as opposed to some other cause such as not noticing the details or feeling falsely safe.
This explores whether 'just leaving it as is' explains why people editing chatbot transcripts by hand kept more personal information in them than people given AI suggestions. The short answer is that the corpus documents the gap but doesn't directly test inaction as the cause. In a study of 188 participants, people who edited manually felt just as confident about their privacy as people who got AI redaction suggestions, yet their final transcripts held substantially more identifiable information Why do people feel safe sharing when they haven't actually removed sensitive information?. The study puts the comfort down to the editing interface itself, not to how much data was actually exposed. That fits a default-inaction story, but it doesn't prove one.
A rival explanation is worth taking seriously: people may not leave the information in on purpose, they may simply not see it. In a separate study of 40 people, more than 30 never noticed that their LLM's memory held, or had inferred, private information about them until someone pointed it out. Once they were shown, every one of them wanted controls to edit and delete it Do users notice when LLMs infer and retain private information?. If personal details are hard to spot, then 'inaction' is really a perception problem. The AI suggestions may help less by nudging people to act than by making the sensitive bits visible in the first place.
A third piece explains why manual editors felt safe anyway. Research on AI-written text finds that having more control over the output raises people's sense of ownership Does user control over AI text shape feelings of ownership?. Free editing feels like control, and that feeling may be standing in for confidence even when little was actually changed. This is also why designs that split the work, with the user stating what they want kept private and the AI carrying out the edits, are proposed as a way to cut the mental effort of privacy management Does splitting privacy work between users and AI reduce cognitive burden?. That approach treats leftover information as a problem of effort and attention, not of motivation.
The broader setting makes leftovers likely in the first place. Talking to a machine with no human judging you encourages deeper self-disclosure How do people decide what to share with AI systems?, so transcripts start out rich in personal detail. Models also volunteer private data inside their own reasoning, where it seems to act as working material they lean on to think Do reasoning traces actually expose private user data?. Much of what needs removing was never consciously 'shared' by the user.
To settle the question you'd need a study that separates three things: whether people saw the sensitive item, whether they decided to remove it, and whether they actually did. The corpus doesn't have that yet. The surprising takeaway is that feeling in control and actually being protected can come apart completely, and the editing interface, not the person, may be what sets which one you get.
Sources 6 notes
Across 188 participants, those manually editing chatbot transcripts reported equivalent privacy confidence as those using AI suggestions, yet left substantially more identifiable information in their final transcripts. This decoupling suggests comfort reflects editing interface design rather than actual data exposure.
A 40-participant study found 30+ users never noticed their LLM's memory feature or inferred private information within it. Once shown the risk, all participants wanted editing and deletion controls, citing privacy as the primary concern.
Study 1 found that greater user control over generated text raised sense of ownership, while personalizing the AI model had no impact on the AI Ghostwriter Effect.
MemoAnalyzer's design divides memory-privacy labor: users state preferences, AI executes edits. This role split is proposed as a way to reduce cognitive load and improve perceived agency, though validation metrics and conflict-resolution processes remain undescribed.
Conversational AI creates a paradoxical disclosure environment where the lack of human judgment simultaneously facilitates intimate self-disclosure (users reciprocate emotional sharing) and incentivizes deception (people self-select toward machines to avoid the psychological cost of lying to humans).
Show all 6 sources
74.8% of privacy leaks in language model reasoning traces result from models materializing sensitive user data during thought processes. Longer reasoning chains amplify leakage, and anonymizing traces post-hoc degrades model utility, suggesting private data functions as cognitive scaffolding.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- "Ghost of the past": identifying and resolving privacy leakage from LLM's memory through proactive user interaction
- Privacy Leakage Overshadowed by Views of AI: A Study on Human Oversight of Privacy in Language Model Agent
- Understanding Users' Privacy Perceptions Towards LLM's RAG-based Memory
- Leaky Thoughts: Large Reasoning Models Are Not Private Thinkers
- Assessing and Mitigating Data Memorization Risks in Fine-Tuned Large Language Models
- The Algorithmic Self-Portrait: Deconstructing Memory in ChatGPT
- Disclose with Care: Designing Privacy Controls in Interview Chatbots
- The Impact of Security and Privacy Controls on Users' Emotional Engagement with Generative AI Chatbots