In Curtin University's controlled experiment, participants reported a similar willingness to disclose to AI and human interviewers. They also gave comparable ratings for trustworthiness, positive experience, awkwardness and their ability to answer effectively. Human interviewers created stronger connection, but rapport did not significantly predict willingness to disclose in the study's regression model.
This is evidence of comparable self-reported disclosure, not proof that the amount, honesty or analytical quality of the information was identical. Those content-level outcomes were not directly compared.
Curtin randomly assigned 60 English-proficient university students and staff to 32 AI-moderated or 28 human-moderated interviews. The topic was fast fashion, including the tension between personal purchasing and the industry's negative effects.
After each interview, participants completed established scales covering willingness to disclose, ability to disclose effectively, trust, connection, awkwardness and the overall experience. The study also recorded facial expressions, skin conductance and heart rate.
The human interviewer followed questions and follow-ups generated by the AI system. This controlled question content across conditions but did not compare AI with a human moderator who was free to probe adaptively.
No statistically significant difference appeared.
These results indicate that participants did not report being less willing or less able to share in the AI interview.
Curtin's regression model explained 57% of variation in willingness to disclose. Two significant predictors were:
Sense of connection was not a significant predictor in this model. Neither joy nor physiological stress significantly predicted disclosure. The evidence therefore separates rapport from disclosure. Human connection was stronger, but trust and the quality of the experience were more closely associated with willingness to share.
No.
A participant can say they were willing to disclose without necessarily providing more detail, greater accuracy or more sensitive information. Curtin did not report a blinded comparison of transcript length, factual specificity or thematic depth between its AI and human conditions.
The most precise claim is that AI matched the scripted human condition on perceived willingness and ability to disclose.
Curtin found no significant difference in trustworthiness: 5.77 for the human interviewer and 5.41 for AI, p = .21. Responsive Research also found that participants were comfortable sharing personal and sensitive experiences in AI-moderated menopause interviews.
Mannheim provides a separate comparison with surveys. Participants trusted the AI format more than the static questionnaire, 4.45 versus 3.95. That result supports the acceptability of AI but does not compare it with a person.
The five studies do not directly test honesty. They do not verify disclosures against external facts or isolate fear of judgment as a causal mechanism.
Curtin found that human presence increased connection and positive engagement without increasing self-reported willingness to disclose. This is compatible with the idea that human warmth and disclosure are distinct, but it cannot establish that AI makes people more honest.
There is no direct social-desirability experiment in the five papers.
Curtin chose a topic involving moral tension, but it did not compare answers with a known truth or a validated social-desirability scale. Responsive Research covered menopause, a sensitive health context, and participants reported comfort. These findings show feasibility, not a measured reduction in bias.
For willingness to disclose in Curtin's model, trust was a significant predictor and connection was not. That supports trust as the more relevant measured factor in this experiment. It should not be generalized into a universal rule. Rapport may affect narrative depth, emotional safety or willingness to revisit a painful memory, outcomes that Curtin did not test at content level.
Human Highway compared text and voice on response quality and participant experience, not on sensitive disclosure. Voice responses were longer and more developed, but participants self-selected their mode.
Longer spoken answers should not be interpreted as proof that voice produces more honest or more sensitive disclosure. The five studies do not answer that question.
Curtin explicitly assigned participants to AI or human interviewer conditions, so they were aware of the difference. The study did not include a deceptive or blinded condition in which the same agent was described differently. It therefore shows disclosure under transparent AI use, but it cannot quantify the effect of disclosure itself.
The current evidence is limited.
Mannheim included adults aged 18 to 55 and found its main format effects remained directionally consistent after controlling for age and gender. Curtin recruited university students and staff. Nottingham recruited a UK sample representative by age, gender and region, but did not compare disclosure across those groups. Human Highway weighted its samples by demographic factors.
None of the studies provides a cross-cultural disclosure comparison or a dedicated analysis of digital confidence.
The studies suggest a plausible use, but do not validate those populations. Responsive Research studied menopause, and Curtin studied personal justifications around fast fashion. Mannheim discusses possible B2B and employee applications, but its sample was consumer panelists. No study directly compares employee fear of retaliation or patient disclosure in a clinical setting.
Claims for employee and patient research should be qualified until those contexts are tested.
Responsive Research sessions averaged about 24 minutes, and Curtin sessions about 16 minutes. Neither study reports a time-series analysis of disclosure or a threshold at which openness declines.
Human Highway observed fewer qualitative signs of cognitive fatigue with conversational AI than with the traditional questionnaire. That does not establish how long an AI interview can run before fatigue affects disclosure.
The five papers do not test crisis protocols, automated risk detection or human escalation. Curtin explicitly recommends human moderation when a topic requires intensive emotional attunement and notes that trauma or sensitive health research needs further study.
For high-risk topics, the current evidence does not support relying on AI alone. Researchers need a separate safeguarding design and should not cite these studies as validation of crisis handling.
The studies used transparent AI conditions, but they do not experimentally compare disclosure with and without explicit disclosure of the moderator's identity. They therefore support the feasibility of informed participation with AI. They do not establish a complete consent standard.
AI can support disclosure when participants trust the interviewer and have a positive experience. Lower human-like rapport does not automatically suppress willingness to share.
Researchers should still distinguish four separate questions:
Only the first of these has a direct AI-versus-human result in the current Glaut Research evidence.
No significant reduction appeared in Curtin's sample.
Yes. Connection was about 26% higher with the human interviewer.
No significant effect appeared in Curtin's regression model.
No. Honesty was not independently verified.
No. The reviewed studies do not test those safeguards.
Collect, analyze, and report research from any source with more depth, speed, and control.
Schedule a free demo
