559 adults read a health recommendation, with a voice reading it aloud, from a doctor, a standard AI or a humanlike AI. Then everyone rated all three recommenders in a random order. That gives two tests of the same idea: one when people judged a single recommender, and one when they compared. This page gives the main findings and what they mean. Every table and figure is on the next page, rebuilt from the released data.
559 adults first read and rated one recommender: a doctor, a standard AI or a humanlike AI, each with a voice reading the advice. The humanlike AI scored 0.94 on trust and the standard AI 0.75.
The difference is small (d = 0.13, p = .188). The data are 2.5 times more likely under no difference, which is anecdotal, not firm evidence.
d = 0.13
humanlike AI against standard AI, each rated alone
Study 4, Part 1.
Everyone then rated all three recommenders. In this comparison the humanlike AI was trusted 0.5 points more than the standard AI, 1.02 against 0.51 (d = 0.36, p < .001).
192 of 504 people trusted the humanlike AI more, 251 trusted both the same, and 61 trusted it less.
+0.5 points
more trust in the humanlike AI than in the standard AI, among the same people
Trust runs from −3 to 3. 504 people in Part 2.
The doctor was trusted 0.6 points more than the standard AI, but only 0.11 more than the humanlike AI (1.13 against 1.02). The gap in effect size shrank from d = 0.41 to 0.09.
The data are only 1.3 times more likely under no difference between the doctor and the humanlike AI, so the gap is small, not proven absent.
Compared with the standard AI Compared with the humanlike AI
Doctor advantage in trust (d)
Doctor minus AI, side by side. Study 4, Part 2.
The same humanlike-versus-standard contrast was d = 0.13 when each person saw one, and d = 0.36 when they saw all three. For belief it went from 0.15 to 0.32.
Part 1 and Part 2 use different formulas for d, so I read the pattern, not the exact ratio. The shift is large enough to matter either way.
Seen alone Side by side
Trust
Belief
Humanlike AI against standard AI, effect size (d).
Side by side, the humanlike AI was rated warmer (d = 0.50) and more competent (d = 0.46) than the standard AI. When I put warmth and competence into the model, the humanlike advantage over the standard AI went from 0.50 to 0.02 points.
That pattern fits the idea that people trusted the humanlike AI more because it seemed warmer and more capable.
0.50 to 0.02
humanlike advantage over the standard AI in trust points, before and after accounting for warmth and competence
Study 4, Part 2 mixed model.
Humanlike framing helps through comparison. With one recommender at a time, the effect was small and uncertain, as in Study 3. When people could weigh the three against each other, the humanlike AI came out clearly ahead of the standard one. A person judging an AI on its own has little to anchor on. Next to alternatives, the cues in the humanlike version become visible.
The effect is about impressions of warmth and competence. The humanlike AI looked warmer and more competent, and those ratings accounted for the whole advantage in the model. This is a statistical account, not proof of mechanism, but it points to what to measure in a real product: how capable and warm an AI seems to its users.
The check worked this time. People rated the humanlike AI as more humanlike than the standard one (d = 0.28, p = .004), which Study 3 had not shown. That supports reading the Part 1 result as a weak effect, not a failed manipulation.
A correction to the dissertation. The dissertation reports the omnibus repeated-measures test of trust as F(2, 1573) = 1.65, p = .192. I could not reproduce that. A repeated-measures ANOVA on the 493 people who rated all three gives F(2, 984) = 55.54, p < .001, which agrees with the pairwise tests. The pairwise results and every effect size elsewhere are unaffected.
Cautions. Part 2 reuses the people from Part 1, and everyone who rated all three had already formed a first impression. The order was randomized, but I did not test order effects here. The advice was a single scenario, and trust is stated trust, not behavior.
Everything on these pages was computed in Python from the released file. Part 1 uses Welch t tests and one-way ANOVAs with Tukey adjustment, with d from the pooled standard deviation. Part 2 uses paired t tests with a Bonferroni correction, dz as the effect size, a repeated-measures ANOVA with the Greenhouse-Geisser correction, and mixed models with a random intercept for each person. The released file has 559 people and shuffled row order, with no IDs, locations, timestamps or free text, and age in four bands.
The dissertation reports Cohen's d as 2t divided by the square root of the degrees of freedom. These pages use the pooled standard deviation for Part 1, which gives somewhat different values from the dissertation's for some tests. The dissertation also lists a sample of 556 people where the released file has 559; I use the file.
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.