854 adults rated a doctor, a humanlike AI and a standard AI on the same health recommendation. Some also heard it spoken, and some only read it. They rated trust, belief, warmth, competence and how much of a mind each seemed to have. This page gives the main findings and what they mean. Every table and figure is on the next page, rebuilt from the released data.
854 adults rated a doctor, a humanlike AI and a standard AI on trust and belief, from −3 to 3. The averages were 1.26 for the doctor, 1.14 for the humanlike AI and 0.47 for the standard AI.
Asked to rank the three on trust, 452 of 721 people put the doctor first. Both orderings point the same way.
63%
of people ranked the doctor first for trust
452 of 721 people in Study 5. Chance would be 33%.
Rated side by side, the humanlike AI was trusted 0.67 points more than the standard AI (d = 0.46, p < .001). That is about as large as the doctor's own edge over the standard AI, 0.80 points (d = 0.50).
341 of 727 people trusted the humanlike AI more, and 94 trusted it less.
Humanlike AI Standard AI
Warmth (0 to 4)
Competence (0 to 4)
Mind perception (0 to 10)
Average ratings, Study 5.
The doctor was trusted 0.13 points more than the humanlike AI (d = 0.13, p < .001). The gap is small, but it is clear in this sample: the data are 75 times more likely under a difference than under none, very strong evidence.
So a humanlike AI closed most of the gap to the doctor without closing all of it.
Compared with the standard AI Compared with the humanlike AI
Doctor advantage in trust and belief (d)
Doctor minus AI, side by side. Study 5.
Half of the people also heard the recommendation spoken. Overall, the voice did not change ratings (the data are 4.6 times more likely under no effect, moderate evidence).
There were hints that a voice helped the humanlike AI (+0.25, p = .017) and hurt the standard AI (−0.26, p = .036). Neither is firm. For the humanlike AI the data are only 2 times more likely under an effect, and for the standard AI they are about equally likely under an effect.
d = 0.05
voiced against written only, all three recommenders averaged
Zero is no difference. Study 5.
People credited the doctor with more of a mind than the standard AI (6.4 against 5.4 out of 10), and credited the humanlike AI with more too. Mind perception was the average of seven items, such as whether the recommender could understand or anticipate.
It carried 57% of the doctor's edge over the standard AI and 59% of the humanlike AI's. In the humanlike-versus-doctor gap, mind perception carried almost none of it.
59%
of the humanlike AI's advantage over the standard AI runs through perceived mind
Mediation analysis, Study 5.
Making an AI seem human can raise trust, if people can compare. As in Study 4, when people judge one recommender the framing barely moves trust, and when they judge several it does. The effect is about as large as the doctor's advantage, so a modest change in how an AI is presented can move trust a long way.
Perceived mind is the main route. About 59% of the humanlike AI's gain over the standard AI ran through mind perception. This is statistical mediation in a design where the recommender was manipulated and mind perception was measured, so it supports the account without proving the order of cause. It fits Chapter 2, where how capable people thought a recommender was tracked trust.
A voice is not the lever. The overall test of voice was null, with moderate evidence for no effect. The recommender-by-voice interaction was significant (F(2, 1450) = 15.89, partial η² = .021), because the voiced humanlike AI gained and the voiced standard AI lost, but each of those results is close to the edge of evidence and was not preregistered per recommender.
A caution for design and policy. The same lever works on a weak AI. If a humanlike presentation raises trust by raising perceived mind, it can make a poor system look better than it is. That is a reason to pair any such design with honest information about what the system can do.
A correction to the dissertation. The dissertation's comparison of the humanlike AI with the doctor has its sign reversed relative to the data. In the data the doctor was rated higher (dz = -0.13 for humanlike minus doctor). I have used the direction the data show.
Cautions. Trust and belief were averaged into one rating because they correlate at r = .81. The recommendation was one scenario, and ratings are stated trust, not behavior.
Everything on these pages was computed in Python from the released file. Paired comparisons are t tests with a Bonferroni correction and dz as the effect size. Voice comparisons are Welch t tests. The mixed ANOVA is on the 727 people who rated all three recommenders, and the mixed models use a random intercept for each person. Mediation uses the Montoya and Hayes difference-score method for repeated measures, with 3,000 bootstrap resamples of people. The released file has 854 people with shuffled row order, no IDs, locations, timestamps or free text, and age in four bands.
The dissertation reports 596 people where the released file has 854, and its mixed-ANOVA degrees of freedom imply a larger sample than the 727 who rated all three. The recommender effect and the interaction agree closely with the dissertation; I use the file throughout.
The data, codebook and code are shared under CC BY 4.0 (data) and the MIT license (code). Please do not use the file to try to identify anyone. Please cite the dissertation: Leitao, M. R. (2024). Artificial Intelligence, Humanization, and Health Beliefs. Georgetown University.