Does making an AI sound human increase user trust?
Yes, and that's a risk: a voice and first-person “I” raised how accurate people thought an LLM's answers were, and human-like chat encouraged more sensitive disclosures.
One large experiment with a pseudo-LLM plus a qualitative study of real ChatGPT use.
experiment, qualitative study2 of 2 tested with LLMsEvidence from 2024How we rateUse human-like cues deliberately, knowing they raise trust whether or not answers are right.
- Treat voice and “I” phrasing as trust boosters that need matching reliability.From the study
- Remind users what happens to their data when the interface feels personal.From the study
- Avoid persona cues in high-risk advice unless accuracy is verified.Our reading
- Test persona choices on trust calibration, not just satisfaction.Our reading
What they found
Cohn et al. (2024)experiment
Online experiment with 2,165 US participants (aged 18–90) using a pseudo-LLM that varied modality (text only vs. speech + text) and grammatical person (“I” vs. “the system”). Speech + text raised anthropomorphism and ratings of information accuracy. First-person “I” raised accuracy ratings and lowered risk ratings, but only in one context.
“the “speech + text” condition led to higher anthropomorphism of the system overall, as well as higher ratings of accuracy of the information the system provides”
Cohn et al., 2024 · Abstract
Zhang et al. (2024)qualitative study
Analysed 200 real ChatGPT sessions (10,380 messages) from the ShareGPT dataset and interviewed 19 users. People disclosed many kinds of personal information, traded privacy for utility and convenience, held wrong mental models of how their data is used, and most hadn't heard of the opt-out from training. Human-like interactions encouraged more sensitive disclosures.
“the human-like interactions encouraged more sensitive disclosures, which complicated users’ ability to navigate the trade-offs”
Zhang et al., 2024 · Abstract
Trade-offs and when it doesn't apply
- The first-person effect appeared in only one context.
- A pseudo-LLM with scripted answers was used.
- Human-like design can improve engagement and accessibility.
What it looks like
How to check it in your product
Compare trust and error-detection between persona and neutral variants on seeded wrong answers.
Open questions
- Does anthropomorphism raise over-reliance on wrong answers specifically?
- How do these effects differ across ages and cultures?
- Can a persona be warm without inflating perceived accuracy?
FAQ
Does a voice make users trust AI answers more?
In Cohn et al. (2,165 participants), speech plus text raised anthropomorphism and perceived accuracy compared with text only.
Do human-like chatbots make people share more?
Zhang et al. found human-like interactions encouraged more sensitive disclosures.
Sources
- Believing Anthropomorphism: Examining the Role of Anthropomorphic Cues on Trust in Large Language ModelsMichelle Cohn, Mahima Pushkarna, Gbolahan O. Olanubi, et al. (2024). CHI 2024 (author version, arXiv:2405.06079).
- “It's a Fair Game”, or Is It? Examining How Users Navigate Disclosure Risks and Benefits When Using LLM-Based Conversational AgentsZhiping Zhang, Michelle Jia, Hao-Ping (Hank) Lee, et al. (2024). CHI 2024.