Can AI suggestions help people remove their own personal information?

A between-subjects study of 188 participants, built around a prototype interview chatbot called EDIT (Enhanced Disclosure Intervention Tool), compared three post-interview privacy conditions: no editing, free editing, and AI-aided editing, where the system "scans the entire interview log for personally identifiable information (PII), and generates suggestions to either redact the PII to a type placeholder... or abstract them to a more general level." The study "confirm[ed] the prevalent issue of oversharing in chatbot-based interviews" and found that AI-aided editing "serves as an effective privacy-control mechanism, reducing PII disclosure while maintaining data quality and user engagement." Final transcripts in the AI-aided condition "contained far fewer residual PII tokens than both free-editing and control."

The researchers attribute the effect to two mechanisms working together: heightened awareness, where "AI suggestions bring to the fore specific spans that might seem harmless when disclosed individually but appear risky when considered together... prompting removals or abstractions that manual passes often miss," and reduced effort, where accepting a one-click replace-or-abstract suggestion is cheaper than manually finding and rewriting disclosive text. They frame free editing, by contrast, as leaving the default of an unedited transcript largely intact, since "prior work on privacy apathy suggests that many people will simply stick with such a status quo even when given a manual editing option" — so AI suggestions "'unlock' more PII reductions without requiring a corresponding rise in privacy concern."

This sits alongside Does chatbot personalization build trust or expose privacy risks?, which shows privacy risk climbing as a side effect of trust and personalization over repeated interaction; this study instead shows a post-hoc design response — it doesn't reduce perceived privacy risk, just the objective PII a transcript carries afterward. It also runs in parallel with Can a smaller user model subset match full model performance?, where a system bounds what it discloses about a user at retrieval time; here the system instead bounds what a user has already disclosed about themselves, after the fact. See also Why do people feel safe sharing when they haven't actually removed sensitive information? for why self-reported comfort tracked none of this.

The study doesn't establish that AI-aided editing changes disclosure during the interview itself — only afterward; its effect could shrink if participants come to expect and discount later editing, or if questions solicit less identifiable detail to begin with. It is also one chatbot, one sensitive topic, and one editing interface, so generalizing to other interview protocols, topics, or the "both-edit" dark-pattern risk the authors flag — where some participants add detail back in after accepting an abstraction — remains untested.