Study 3: every table and figure

Each analysis from the dissertation, rebuilt from the released data, with a figure beside every table. Hover over any mark for the details.

Deep dive: every table and figure

The rest of this page rebuilds each analysis from the public data. Hover over any mark for the details, and open the guide below if a term is unfamiliar.

Plain-language guide to the terms

Trust: How willing someone is to follow the advice. In these studies, people rated this from 0 (not at all) to 10 (completely).
Quantitative reasoning: Deciding with numbers, data and statistics, like reading a lab result.
Intuition: Deciding with experience and judgment, like a doctor sensing that something is off when the tests look fine.
Effect size (Cohen's d): How big a difference is, on a scale that can be compared across studies. This site uses the Funder and Ozer (2019) labels for psychology: about 0.20 is small, 0.41 is medium, 0.63 is large, and 0.87 or more is very large.
95% interval: The range of values the true average probably falls in. A narrow range means a precise estimate; a range that includes zero means the difference may not be real.
p value: How surprising the result would be if there were no real difference. A p below .05 is the usual bar for calling a result reliable, and "p < .001" means chance is a very unlikely explanation.
Mixed model: A way to analyze data when each person gives many ratings, so that people who rate everything high or low do not distort the results.
Mediation: A test of whether one thing explains part of the link between two others. Here: does seeing a doctor as more competent explain part of why people trust the doctor more?
Preregistered: The plan for the study and its analysis was written and posted publicly before anyone looked at the data.

How to read this page. Each person saw one source only, so every comparison is between groups of different people. Trust and belief are on a scale from −3 to 3, warmth and competence from 0 to 4. The public file groups age into four bands, so the models that used age in years use the band number (1 = 18 to 44, 4 = 70 and over); their age coefficients differ from the dissertation's in scale.

Who took part

609 adults rated trust. 240 saw the humanlike ai, 240 saw the standard ai, 129 saw the doctor.

Table 1. The people in the released file

What to look for. The share in each group. Younger adults and women make up the largest groups.

Group	People	Share
Age group: 18-44	312	51.2%
Age group: 45-59	130	21.3%
Age group: 60-69	61	10.0%
Age group: 70+	55	9.0%
Gender: Woman	315	51.7%
Gender: Man	247	40.6%
Race: White	356	58.5%
Race: Another or multiple	212	34.8%
Education: No college degree	365	59.9%
Education: College degree or more	203	33.3%
How religious: Not at all to slightly	226	37.1%
How religious: Moderately	164	26.9%
How religious: More than moderately	177	29.1%

Shares are of all 609 people. Left blank in the file: age group (51), gender (47), race (41), education (41), how religious (42). Race is shown as White or another or multiple because no other group reached 30 people. Age is in bands; exact age is not released.

Table 1 as a figure

Each bar is the share of the 609 people in that group. Hover a bar for the head count.

The three sources and the check

People were randomly assigned to read the same health recommendation from a doctor, from a standard AI or from an AI described in humanlike terms. To check the framing, people in the two AI conditions rated how humanlike the AI seemed on a scale from 0 to 4.

Table 2. How humanlike the AI seemed (0 to 4)

What to look for. The two means are close together and the interval for the difference includes zero, so people did not see the humanlike AI as more humanlike.

Comparison	Humanlike AI, Standard AI: mean (SD)	Difference [95% CI]	t (df)	p	d [95% CI]
How humanlike the AI seemed: humanlike AI versus standard AI	1.64 (1.16), 1.54 (1.04)	0.10 [−0.10, 0.30]	0.98 (471)	.328	0.09 [−0.09, 0.27]

Welch t test, two-sided. Only the two AI conditions answered this question.

Figure 1. How humanlike the AI seemed

Each dot is one person's rating, from 0 (not at all humanlike) to 4 (very humanlike). The two clouds overlap almost completely.

Outline: how answers are spread (smoothed). Dots: one person each. Diamond and bar: the average and its bootstrapped 95% interval.

View as a table

Measure	n	Mean	95% interval
Humanlike AI	239	1.6	1.5 to 1.8
Standard AI	240	1.5	1.4 to 1.7

Trust, belief, warmth and competence

The preregistered questions were whether the doctor was trusted more than the AIs, and whether the humanlike AI was trusted more than the standard one. The tables give the descriptives, the omnibus test and then each comparison.

Table 3. Ratings by source

What to look for. Compare means within a measure. The doctor is highest on trust, belief and warmth; the two AIs sit close together.

Measure and source	n	Mean	SD	SE	95% CI	Median	Min	Max
Trust (−3 to 3): Humanlike AI	240	0.52	1.46	0.09	[0.33, 0.70]	1.00	−3.00	3.00
Trust (−3 to 3): Standard AI	240	0.41	1.53	0.10	[0.22, 0.61]	1.00	−3.00	3.00
Trust (−3 to 3): Doctor	129	0.86	1.59	0.14	[0.58, 1.14]	1.00	−3.00	3.00
Belief (−3 to 3): Humanlike AI	240	0.65	1.49	0.10	[0.47, 0.84]	1.00	−3.00	3.00
Belief (−3 to 3): Standard AI	240	0.61	1.51	0.10	[0.42, 0.80]	1.00	−3.00	3.00
Belief (−3 to 3): Doctor	129	0.98	1.41	0.12	[0.74, 1.23]	1.00	−3.00	3.00
Warmth (0 to 4): Humanlike AI	239	1.54	1.15	0.07	[1.39, 1.69]	2.00	0.00	4.00
Warmth (0 to 4): Standard AI	240	1.44	1.18	0.08	[1.29, 1.59]	1.00	0.00	4.00
Warmth (0 to 4): Doctor	128	2.00	1.08	0.10	[1.81, 2.19]	2.00	0.00	4.00
Competence (0 to 4): Humanlike AI	240	1.86	1.12	0.07	[1.72, 2.00]	2.00	0.00	4.00
Competence (0 to 4): Standard AI	239	2.00	1.09	0.07	[1.86, 2.13]	2.00	0.00	4.00
Competence (0 to 4): Doctor	129	2.25	0.96	0.08	[2.08, 2.42]	2.00	0.00	4.00

Standard errors and intervals are for each group's mean.

Figure 2. Trust, every person

Each dot is one person's trust rating, from −3 (no trust) to 3 (complete trust). The curve shows where ratings pile up, and the bar marks the mean with its 95% interval.

Outline: how answers are spread (smoothed). Dots: one person each. Diamond and bar: the average and its bootstrapped 95% interval.

View as a table

Measure	n	Mean	95% interval
Doctor	129	0.9	0.6 to 1.1
Humanlike AI	240	0.5	0.3 to 0.7
Standard AI	240	0.4	0.2 to 0.6

Figure 3. Belief, every person

The same display for belief in the recommendation.

Outline: how answers are spread (smoothed). Dots: one person each. Diamond and bar: the average and its bootstrapped 95% interval.

View as a table

Measure	n	Mean	95% interval
Doctor	129	1.0	0.7 to 1.2
Humanlike AI	240	0.7	0.5 to 0.8
Standard AI	240	0.6	0.4 to 0.8

Figure 4. Warmth, every person

The same display for how warm the source seemed, from 0 to 4.

Outline: how answers are spread (smoothed). Dots: one person each. Diamond and bar: the average and its bootstrapped 95% interval.

View as a table

Measure	n	Mean	95% interval
Doctor	128	2.0	1.8 to 2.2
Humanlike AI	239	1.5	1.4 to 1.7
Standard AI	240	1.4	1.3 to 1.6

Figure 5. Competence, every person

The same display for how competent the source seemed, from 0 to 4.

Outline: how answers are spread (smoothed). Dots: one person each. Diamond and bar: the average and its bootstrapped 95% interval.

View as a table

Measure	n	Mean	95% interval
Doctor	129	2.3	2.1 to 2.4
Humanlike AI	240	1.9	1.7 to 2.0
Standard AI	239	2.0	1.9 to 2.1

Figure 6. Means and intervals

Table 3 as a figure. Each dot is a mean with its 95% interval, one panel per measure. Where the doctor's interval sits clear of the other two, the doctor was rated higher.

View as a table

Source	Trust: mean [95% interval]	Belief: mean [95% interval]	Warmth: mean [95% interval]	Competence: mean [95% interval]
Doctor	0.9 [0.6, 1.1]	1.0 [0.7, 1.2]	2.0 [1.8, 2.2]	2.3 [2.1, 2.4]
Humanlike AI	0.5 [0.3, 0.7]	0.7 [0.5, 0.8]	1.5 [1.4, 1.7]	1.9 [1.7, 2.0]
Standard AI	0.4 [0.2, 0.6]	0.6 [0.4, 0.8]	1.4 [1.3, 1.6]	2.0 [1.9, 2.1]

Table 4. One-way ANOVAs across the three sources

What to look for. The F and p columns say whether the three sources differ at all; the last three columns say which pairs do, after adjusting for three comparisons.

Measure	n	F (df)	p	η²	Tukey p: doctor, standard AI	Tukey p: doctor, humanlike AI	Tukey p: humanlike, standard AI
Trust	609	F(2, 606) = 3.75	.024	.012	.019	.096	.732
Belief	609	F(2, 606) = 2.91	.055	.010	.056	.103	.949
Warmth	607	F(2, 604) = 10.55	< .001	.034	< .001	< .001	.592
Competence	608	F(2, 605) = 5.52	.004	.018	.081	.003	.341

Tukey HSD for the three pairwise comparisons.

Table 4 as a figure

Each bar is η², the share of the variation in that rating that the source explains. All are small; warmth is the largest.

Table 5. Planned and additional comparisons

What to look for. The d column, with its interval, tells you how large each difference is. Compare the doctor rows with the humanlike-versus-standard rows.

Comparison	First, Second: mean (SD)	Difference [95% CI]	t (df)	p	d [95% CI]	Kind	Tukey p
Trust: Doctor versus the two AI conditions combined	0.86 (1.59), 0.46 (1.50)	0.40 [0.09, 0.70]	2.54 (193.21)	.012	0.26 [0.07, 0.46]	Planned
Trust: Humanlike AI versus standard AI	0.52 (1.46), 0.41 (1.53)	0.10 [−0.16, 0.37]	0.76 (477.11)	.446	0.07 [−0.11, 0.25]	Planned	.732
Trust: Doctor versus standard AI	0.86 (1.59), 0.41 (1.53)	0.45 [0.11, 0.79]	2.62 (253.27)	.009	0.29 [0.07, 0.50]	Additional	.019
Trust: Doctor versus humanlike AI	0.86 (1.59), 0.52 (1.46)	0.34 [0.01, 0.68]	2.04 (244.06)	.043	0.23 [0.01, 0.44]	Additional	.096
Belief: Doctor versus the two AI conditions combined	0.98 (1.41), 0.63 (1.50)	0.35 [0.07, 0.63]	2.47 (211.42)	.014	0.24 [0.04, 0.43]	Planned
Belief: Humanlike AI versus standard AI	0.65 (1.49), 0.61 (1.51)	0.04 [−0.23, 0.31]	0.30 (477.91)	.761	0.03 [−0.15, 0.21]	Planned	.949
Belief: Doctor versus standard AI	0.98 (1.41), 0.61 (1.51)	0.37 [0.06, 0.68]	2.35 (276.78)	.019	0.25 [0.04, 0.47]	Additional	.056
Belief: Doctor versus humanlike AI	0.98 (1.41), 0.65 (1.49)	0.33 [0.02, 0.64]	2.10 (273.56)	.037	0.23 [0.01, 0.44]	Additional	.103
Warmth: Doctor versus the two AI conditions combined	2.00 (1.08), 1.49 (1.16)	0.51 [0.30, 0.73]	4.68 (212.45)	< .001	0.45 [0.25, 0.64]	Planned
Warmth: Humanlike AI versus standard AI	1.54 (1.15), 1.44 (1.18)	0.10 [−0.11, 0.31]	0.96 (476.79)	.336	0.09 [−0.09, 0.27]	Planned	.592
Warmth: Doctor versus standard AI	2.00 (1.08), 1.44 (1.18)	0.56 [0.32, 0.80]	4.61 (279.4)	< .001	0.49 [0.27, 0.71]	Additional	< .001
Warmth: Doctor versus humanlike AI	2.00 (1.08), 1.54 (1.15)	0.46 [0.22, 0.70]	3.81 (273.72)	< .001	0.41 [0.19, 0.63]	Additional	< .001
Competence: Doctor versus the two AI conditions combined	2.25 (0.96), 1.93 (1.10)	0.32 [0.13, 0.52]	3.26 (227.57)	.001	0.30 [0.10, 0.49]	Planned
Competence: Humanlike AI versus standard AI	1.86 (1.12), 2.00 (1.09)	−0.14 [−0.34, 0.06]	−1.36 (476.84)	.173	−0.12 [−0.30, 0.05]	Planned	.341
Competence: Doctor versus standard AI	2.25 (0.96), 2.00 (1.09)	0.25 [0.04, 0.47]	2.29 (292.05)	.023	0.24 [0.03, 0.46]	Additional	.081
Competence: Doctor versus humanlike AI	2.25 (0.96), 1.86 (1.12)	0.39 [0.17, 0.61]	3.51 (297.15)	< .001	0.37 [0.15, 0.58]	Additional	.003

Welch t tests, two-sided. d uses the pooled standard deviation and is positive when the first group is higher. Planned comparisons were preregistered; the rest are additional. The Tukey column adjusts for the three pairwise comparisons.

Figure 7. Effect sizes

Table 5 as a figure. Each dot is Cohen's d with a 95% interval. The dashed lines mark a small, medium and large effect on the Funder and Ozer scale. An interval that crosses zero cannot rule out no difference.

Table 6. What the study could detect

What to look for. The d column is the smallest true difference this design would find 80% of the time. Smaller true differences could easily have been missed.

Comparison	n, first	n, second	Smallest detectable d
Doctor versus the two AI conditions combined	129	480	0.28
Humanlike AI versus standard AI	240	240	0.26
Doctor versus standard AI	129	240	0.31
Doctor versus humanlike AI	129	240	0.31

Two-sided Welch test, α = .05, power = .80.

Table 6 as a figure

Each bar is the smallest effect, in standard deviations, that the comparison could reliably detect.

Regression models

Each model predicts a rating from the source. The reference group is the humanlike AI. The second model adds gender and age group, and the third adds warmth and competence. When warmth and competence enter, the doctor's coefficient falls close to zero.

Table 7a. Trust: condition only