Do Participants Overshare With AI Interviewers? 2026 Study
Qualitati Research Team · 2026-08-12 · 7 min read
Participants do overshare with AI interviewers — and a 2026 study shows the fix is not to restrain them during the interview, but to let them redact afterwards. Li, Xiao, and Li tested three post-interview privacy controls with 188 participants and found AI-assisted redaction cut personally identifiable information by 41.2% with no measurable loss in data quality.
What did the study test?
The study asked whether participants can be given practical control over their own transcripts after a chatbot interview ends. According to Li, Xiao, and Li (2026) in "Disclose with Care: Designing Privacy Controls in Interview Chatbots" (arXiv, February 2026; a related version appears in the CHI 2026 Extended Abstracts), the premise is a tension every AI-moderated study inherits: a chatbot feels anonymous and non-judgmental, so people talk freely — and then reveal names, employers, and identifying circumstances they would not have put in a survey box.
The design was between-subjects, with 188 participants (200 recruited, 11 low-quality responses excluded) split across three conditions:
- Control (n = 66) — the interview ends, no chance to edit.
- Free-editing (n = 62) — participants may freely revise their own answers before submitting.
- AI-aided editing (n = 60) — an LLM flags likely identifying spans and suggests replacements or abstractions, which the participant accepts or ignores.
The interview topic was the use of AI in job interviews — moderately sensitive, covering education, current employment, nervous experiences with AI screening, and whether the participant had ever concealed AI use. Six question groups in all.
How much personal information did participants reveal?
Substantially less in the AI-aided condition, and the effect was large. The authors report mean PII per message and a net reduction rate relative to the participant's original text:
| Condition | PII per message (M) | Net PII reduction | Participants who actively reduced PII |
|---|
| AI-aided editing (n = 60) | 0.107 | 41.2% | 71.2% |
| Free-editing (n = 62) | 0.259 | −2.6% (slight increase) | 22.6% |
| Control (n = 66) | 0.247 | — | — |
The between-condition difference was significant: F(2,185) = 30.747, p < .001, with an effect size of η² = 0.249 — roughly a quarter of the variance in disclosure explained by which privacy control participants were given.
The most instructive number is the −2.6%. Given a blank licence to edit, participants did not meaningfully self-redact. Of the edits made in the free-editing condition, 19.23% reduced PII, 11.54% increased it, and 69.23% left identifiability unchanged. Under AI assistance, 95.48% of edits reduced PII. Offering the option is not the intervention; showing people what is identifying is.
Does redacting a transcript damage the data?
No — that is the finding researchers should take seriously. The authors scored answers on a Response Quality Index combining relevance, clarity, and specificity, and found no significant difference across conditions: free-editing M = 5.24 (95% CI [5.06, 5.42]), AI-aided M = 4.95 (95% CI [4.80, 5.10]), control M = 4.92 (95% CI [4.77, 5.08]). None of the three component scores differed either.
Engagement held up too. Median response length was about 14 words in every condition (AI-aided M = 17.5, SD = 15.3; free-editing M = 18.5, SD = 16.4; control M = 17.7, SD = 15.9). And participants' self-reported comfort with disclosure did not differ by condition (F = 1.349, p = 0.262), nor did their willingness to discuss sensitive topics in future (F = 1.602, p = 0.204).
In other words: the privacy gain came almost entirely out of identifiability, not out of substance. Replacing "my manager Priya at [company]" with "my manager at a mid-sized firm" costs an analyst nothing.
Why post-hoc timing matters
The control was applied after the interview by design. Warning participants up front, or interrupting them mid-answer, would suppress the candour that makes an AI-moderated interview worth running in the first place. The authors' recommendation is to let the conversation run naturally, then hand the transcript back for review — capturing the disclosure and then containing it, rather than trading one for the other.
Uptake was voluntary and imperfect: 68.33% of the AI-aided group (41 of 60) used the feature at all, and their edits clustered in the first question group (47.6% of edits) before tapering off. A privacy control that depends on sustained participant attention will decay over a long interview — worth designing around if your protocol runs 30 minutes rather than 10.
What this means for researchers
The paper's design implications translate directly into protocol decisions for anyone running AI-moderated interviews or conversational surveys:
- Make risk visible at the span level. Highlight the specific words that identify someone, in place, rather than issuing a general warning.
- Prefer replace and abstract over delete. Reversible, abstracting edits preserve analytic meaning; deletion destroys it.
- Keep suggestions conservative. Target high-risk entity categories with a domain-tuned taxonomy instead of flagging everything.
- Never force redaction. Acceptance should stay optional — the effect above was achieved with suggestions, not compulsion.
- Place the control after the interview, so naturalistic disclosure is preserved.
- Treat the AI as awareness scaffolding, compensating for privacy-literacy gaps rather than policing participants.
If you run interviews on employment, health, workplace conflict, or anything an employer could act on, this is a cheap addition to your consent architecture — and a stronger ethics-review answer than "transcripts are stored securely". It pairs naturally with de-identifying before analysis: in QualiTaTi, participant-facing collection happens in the AI Interviewer, while coding and theme work happen downstream in ThemeLens, so the redaction step belongs at the boundary between them.
Two limits worth stating. The topic was only moderately sensitive, so the effect on genuinely high-stakes disclosure (immigration status, abuse, illness) is untested. And PII per message is a proxy for identifiability, not a guarantee of anonymity — a quasi-identifier combination can re-identify someone with no names present at all.
FAQ
Do people disclose more to an AI interviewer than to a human?
Prior work suggests they often do, because a chatbot reads as less judgmental — and this study treats that increased disclosure as the starting problem rather than an unambiguous benefit. It did not itself run an AI-versus-human comparison.
Will letting participants edit their transcripts ruin my data?
Not on this evidence. Response quality and length were statistically indistinguishable across all three conditions, and 95.48% of AI-assisted edits reduced identifiability rather than removing content.
Is it enough to just tell participants they can edit their answers?
No. Free editing produced a 2.6% net increase in PII, and only 22.6% of those participants reduced identifying detail. Assistance, not permission, is what changed behaviour.
Should I redact before or after coding?
Before, if the redaction is participant-driven — the point is that identifying detail never enters the analytic corpus. Automated de-identification later still leaves the raw transcript in your system.
Last updated: August 12, 2026.
This is an independent editorial summary of third-party research. QualiTaTi is not affiliated with the authors or their institutions; readers should consult the original paper for full methods and limitations.