Study 5: 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. Everyone rated the doctor, the humanlike AI and the standard AI, so the main comparisons are within people. Voice is the one thing that varies between people: some read the recommendation only, and others also heard it spoken. “Trust and belief” is the average of the two ratings, on a scale from −3 to 3. Warmth and competence run from 0 to 4, mind perception from 0 to 10, and praise and blame from 0 to 100. Age is in four bands in the public file, so the age coefficient differs in scale from the dissertation's.
Who took part
854 adults are in the released file. 727 of them rated all three recommenders, and the paired tests use those who gave both ratings in each comparison.
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 | 369 | 43.2% |
| Age group: 45-59 | 136 | 15.9% |
| Age group: 60-69 | 114 | 13.3% |
| Age group: 70+ | 73 | 8.5% |
| Gender: Woman | 466 | 54.6% |
| Gender: Man | 228 | 26.7% |
| Race: White | 471 | 55.2% |
| Race: Another or multiple | 226 | 26.5% |
| Education: No college degree | 449 | 52.6% |
| Education: College degree or more | 248 | 29.0% |
| How religious: Not at all to slightly | 258 | 30.2% |
| How religious: Moderately | 190 | 22.2% |
| How religious: More than moderately | 248 | 29.0% |
Shares are of all 854 people. Left blank in the file: age group (162), gender (160), race (157), education (157), how religious (158). Race is shown as White or another or multiple because no other group reached 30 people. Age is in bands.
Table 1 as a figure
Each bar is the share of the 854 people in that group. Hover a bar for the head count.
What to look for. Voice is the between-person factor; about six in ten heard the voice.
| Group | All people | Rated all three |
|---|---|---|
| Written only | 349 | 297 |
| Written and voiced | 505 | 430 |
| Total | 854 | 727 |
People were randomly assigned to read the recommendation only or to read it and hear a voice recording.
Ratings of the three recommenders
Every person rated each recommender on trust, belief, warmth, competence, mind perception, how appropriate it was for these decisions, and how much praise or blame it would deserve.
What to look for. Compare means within a measure. The doctor is highest on nearly everything, the humanlike AI is close behind, and the standard AI is lowest.
| Measure and recommender | n | Mean | SD | SE | 95% CI | Median | Min | Max |
|---|---|---|---|---|---|---|---|---|
| Trust/belief (-3 to 3): Doctor | 854 | 1.26 | 1.28 | 0.04 | [1.17, 1.34] | 1.50 | −3.00 | 3.00 |
| Trust/belief (-3 to 3): Humanlike AI | 789 | 1.14 | 1.41 | 0.05 | [1.04, 1.24] | 1.00 | −3.00 | 3.00 |
| Trust/belief (-3 to 3): Standard AI | 727 | 0.47 | 1.69 | 0.06 | [0.35, 0.59] | 0.50 | −3.00 | 3.00 |
| Trust (-3 to 3): Doctor | 854 | 1.18 | 1.43 | 0.05 | [1.09, 1.28] | 1.00 | −3.00 | 3.00 |
| Trust (-3 to 3): Humanlike AI | 789 | 1.06 | 1.54 | 0.05 | [0.95, 1.16] | 1.00 | −3.00 | 3.00 |
| Trust (-3 to 3): Standard AI | 727 | 0.42 | 1.75 | 0.06 | [0.29, 0.55] | 0.00 | −3.00 | 3.00 |
| Belief (-3 to 3): Doctor | 854 | 1.33 | 1.32 | 0.05 | [1.25, 1.42] | 2.00 | −3.00 | 3.00 |
| Belief (-3 to 3): Humanlike AI | 789 | 1.22 | 1.45 | 0.05 | [1.12, 1.32] | 1.00 | −3.00 | 3.00 |
| Belief (-3 to 3): Standard AI | 727 | 0.52 | 1.74 | 0.06 | [0.39, 0.64] | 1.00 | −3.00 | 3.00 |
| Warmth (0 to 4): Doctor | 852 | 2.43 | 1.17 | 0.04 | [2.35, 2.51] | 2.50 | 0.00 | 4.00 |
| Warmth (0 to 4): Humanlike AI | 788 | 2.23 | 1.25 | 0.04 | [2.15, 2.32] | 2.00 | 0.00 | 4.00 |
| Warmth (0 to 4): Standard AI | 726 | 1.61 | 1.40 | 0.05 | [1.51, 1.71] | 2.00 | 0.00 | 4.00 |
| Competence (0 to 4): Doctor | 852 | 2.42 | 1.10 | 0.04 | [2.35, 2.50] | 2.00 | 0.00 | 4.00 |
| Competence (0 to 4): Humanlike AI | 788 | 2.37 | 1.13 | 0.04 | [2.29, 2.44] | 2.00 | 0.00 | 4.00 |
| Competence (0 to 4): Standard AI | 727 | 1.86 | 1.27 | 0.05 | [1.77, 1.95] | 2.00 | 0.00 | 4.00 |
| Mind perception (0 to 10): Doctor | 836 | 6.37 | 2.25 | 0.08 | [6.22, 6.53] | 6.43 | 0.00 | 10.00 |
| Mind perception (0 to 10): Humanlike AI | 767 | 6.31 | 2.38 | 0.09 | [6.14, 6.48] | 6.43 | 0.00 | 10.00 |
| Mind perception (0 to 10): Standard AI | 701 | 5.44 | 2.68 | 0.10 | [5.25, 5.64] | 5.29 | 0.00 | 10.00 |
| Appropriate for these decisions (-3 to 3): Doctor | 854 | 1.12 | 1.45 | 0.05 | [1.03, 1.22] | 1.00 | −3.00 | 3.00 |
| Appropriate for these decisions (-3 to 3): Humanlike AI | 789 | 0.74 | 1.57 | 0.06 | [0.63, 0.85] | 1.00 | −3.00 | 3.00 |
| Appropriate for these decisions (-3 to 3): Standard AI | 727 | 0.12 | 1.80 | 0.07 | [−0.02, 0.25] | 0.00 | −3.00 | 3.00 |
| Praise if it works (0 to 100): Doctor | 852 | 62.81 | 23.14 | 0.79 | [61.26, 64.37] | 63.00 | 0.00 | 100.00 |
| Praise if it works (0 to 100): Humanlike AI | 786 | 62.66 | 24.45 | 0.87 | [60.95, 64.37] | 67.00 | 0.00 | 100.00 |
| Praise if it works (0 to 100): Standard AI | 721 | 62.50 | 25.95 | 0.97 | [60.60, 64.40] | 66.00 | 0.00 | 100.00 |
| Blame if it fails (0 to 100): Doctor | 847 | 62.88 | 25.09 | 0.86 | [61.19, 64.57] | 67.00 | 0.00 | 100.00 |
| Blame if it fails (0 to 100): Humanlike AI | 784 | 62.79 | 26.59 | 0.95 | [60.92, 64.65] | 67.00 | 0.00 | 100.00 |
| Blame if it fails (0 to 100): Standard AI | 723 | 58.33 | 29.33 | 1.09 | [56.19, 60.47] | 60.00 | 0.00 | 100.00 |
Trust and belief are on a scale from −3 to 3. Praise and blame are 0 to 100. The humanlike AI was not rated on being humanlike by the same people.
Figures 1 to 3. Means and intervals
Table 3 as figures. Each dot is a mean with its 95% interval. The first figure shows trust and belief.
View as a table
| Recommender | Trust and belief: mean [95% interval] | Trust: mean [95% interval] | Belief: mean [95% interval] |
|---|---|---|---|
| Doctor | 1.3 [1.2, 1.3] | 1.2 [1.1, 1.3] | 1.3 [1.3, 1.4] |
| Humanlike AI | 1.1 [1.0, 1.2] | 1.1 [0.9, 1.2] | 1.2 [1.1, 1.3] |
| Standard AI | 0.5 [0.3, 0.6] | 0.4 [0.3, 0.6] | 0.5 [0.4, 0.6] |
Warmth, competence and mind perception.
View as a table
| Recommender | Warmth: mean [95% interval] | Competence: mean [95% interval] | Mind perception: mean [95% interval] |
|---|---|---|---|
| Doctor | 2.4 [2.4, 2.5] | 2.4 [2.4, 2.5] | 6.4 [6.2, 6.5] |
| Humanlike AI | 2.2 [2.1, 2.3] | 2.4 [2.3, 2.4] | 6.3 [6.1, 6.5] |
| Standard AI | 1.6 [1.5, 1.7] | 1.9 [1.8, 1.9] | 5.4 [5.3, 5.6] |
How appropriate the recommender was judged to be, and the praise and blame it would get.
View as a table
| Recommender | Appropriate: mean [95% interval] | Praise if it works: mean [95% interval] | Blame if it fails: mean [95% interval] |
|---|---|---|---|
| Doctor | 1.1 [1.0, 1.2] | 62.8 [61.3, 64.4] | 62.9 [61.2, 64.6] |
| Humanlike AI | 0.7 [0.6, 0.8] | 62.7 [61.0, 64.4] | 62.8 [60.9, 64.7] |
| Standard AI | 0.1 [−0.0, 0.3] | 62.5 [60.6, 64.4] | 58.3 [56.2, 60.5] |
Figure 4. Trust and belief, every person
Each dot is one person's average of trust and belief for that recommender, from −3 to 3. The curve shows where ratings pile up, and the bar marks the mean with its bootstrapped 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 | 854 | 1.3 | 1.2 to 1.3 |
| Humanlike AI | 789 | 1.1 | 1.0 to 1.2 |
| Standard AI | 727 | 0.5 | 0.3 to 0.6 |
Figure 5. Mind perception, every person
Each dot is one person's average on the seven mind-perception items, from 0 to 10.
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 | 836 | 6.4 | 6.2 to 6.5 |
| Humanlike AI | 767 | 6.3 | 6.1 to 6.5 |
| Standard AI | 701 | 5.4 | 5.2 to 5.6 |
The three recommenders compared
Each comparison uses the people who rated both recommenders. A positive difference means the first one named was rated higher. The dissertation's central question is how the humanlike AI sits between the standard AI and the doctor.
What to look for. The dz column shows how large each difference is. The humanlike AI is clearly above the standard AI and a little below the doctor.
| Comparison | n | First, second: mean (SD) | Difference [95% CI] | t (df) | p | Bonferroni p | dz [95% CI] | Higher / same / lower |
|---|---|---|---|---|---|---|---|---|
| Trust and belief: Doctor versus standard AI | 727 | 1.27 (1.29), 0.47 (1.69) | 0.80 [0.68, 0.91] | 13.44 (726) | < .001 | < .001 | 0.50 [0.42, 0.58] | 372 / 238 / 117 |
| Trust and belief: Humanlike versus standard AI | 727 | 1.13 (1.41), 0.47 (1.69) | 0.67 [0.56, 0.77] | 12.52 (726) | < .001 | < .001 | 0.46 [0.39, 0.54] | 341 / 292 / 94 |
| Trust and belief: Humanlike AI versus doctor | 789 | 1.14 (1.41), 1.27 (1.28) | −0.13 [−0.20, −0.06] | −3.78 (788) | < .001 | < .001 | −0.13 [−0.20, −0.06] | 184 / 364 / 241 |
| Trust: Doctor versus standard AI | 727 | 1.19 (1.44), 0.42 (1.75) | 0.77 [0.64, 0.90] | 11.80 (726) | < .001 | < .001 | 0.44 [0.36, 0.51] | 331 / 295 / 101 |
| Trust: Humanlike versus standard AI | 727 | 1.05 (1.54), 0.42 (1.75) | 0.63 [0.52, 0.74] | 10.83 (726) | < .001 | < .001 | 0.40 [0.33, 0.48] | 297 / 347 / 83 |
| Trust: Humanlike AI versus doctor | 789 | 1.06 (1.54), 1.20 (1.43) | −0.15 [−0.24, −0.06] | −3.18 (788) | .002 | .005 | −0.11 [−0.18, −0.04] | 131 / 466 / 192 |
| Belief: Doctor versus standard AI | 727 | 1.34 (1.34), 0.52 (1.74) | 0.82 [0.70, 0.95] | 13.03 (726) | < .001 | < .001 | 0.48 [0.41, 0.56] | 340 / 293 / 94 |
| Belief: Humanlike versus standard AI | 727 | 1.22 (1.45), 0.52 (1.74) | 0.70 [0.59, 0.81] | 12.17 (726) | < .001 | < .001 | 0.45 [0.37, 0.53] | 312 / 338 / 77 |
| Belief: Humanlike AI versus doctor | 789 | 1.22 (1.45), 1.34 (1.33) | −0.12 [−0.20, −0.04] | −3.05 (788) | .002 | .007 | −0.11 [−0.18, −0.04] | 147 / 440 / 202 |
| Warmth: Doctor versus standard AI | 725 | 2.44 (1.18), 1.61 (1.40) | 0.83 [0.73, 0.92] | 17.38 (724) | < .001 | < .001 | 0.65 [0.57, 0.73] | 371 / 285 / 69 |
| Warmth: Humanlike versus standard AI | 725 | 2.21 (1.26), 1.61 (1.40) | 0.61 [0.52, 0.69] | 14.02 (724) | < .001 | < .001 | 0.52 [0.44, 0.60] | 309 / 342 / 74 |
| Warmth: Humanlike AI versus doctor | 787 | 2.24 (1.25), 2.44 (1.17) | −0.20 [−0.27, −0.14] | −6.52 (786) | < .001 | < .001 | −0.23 [−0.30, −0.16] | 111 / 447 / 229 |
| Competence: Doctor versus standard AI | 726 | 2.43 (1.11), 1.86 (1.27) | 0.58 [0.49, 0.66] | 13.21 (725) | < .001 | < .001 | 0.49 [0.41, 0.57] | 312 / 325 / 89 |
| Competence: Humanlike versus standard AI | 726 | 2.35 (1.14), 1.86 (1.27) | 0.49 [0.41, 0.57] | 12.57 (725) | < .001 | < .001 | 0.47 [0.39, 0.54] | 299 / 352 / 75 |
| Competence: Humanlike AI versus doctor | 787 | 2.36 (1.12), 2.44 (1.10) | −0.08 [−0.13, −0.02] | −2.62 (786) | .009 | .027 | −0.09 [−0.16, −0.02] | 121 / 503 / 163 |
| Mind perception: Doctor versus standard AI | 688 | 6.40 (2.23), 5.42 (2.69) | 0.98 [0.84, 1.12] | 13.66 (687) | < .001 | < .001 | 0.52 [0.44, 0.60] | 405 / 106 / 177 |
| Mind perception: Humanlike versus standard AI | 683 | 6.29 (2.39), 5.44 (2.67) | 0.85 [0.73, 0.97] | 13.50 (682) | < .001 | < .001 | 0.52 [0.44, 0.60] | 414 / 122 / 147 |
| Mind perception: Humanlike AI versus doctor | 752 | 6.30 (2.40), 6.39 (2.23) | −0.08 [−0.16, −0.01] | −2.15 (751) | .032 | .095 | −0.08 [−0.15, −0.01] | 295 / 168 / 289 |
Paired t tests, two-sided, with a Bonferroni correction across the three comparisons of each measure. dz is positive when the first recommender is rated higher. The humanlike-versus-doctor rows are negative: the doctor was rated higher.
Figure 6. Effect sizes of the paired comparisons
Table 4 as a figure. Each dot is dz with a 95% interval. Negative values mean the first recommender was rated lower. The dashed lines mark a small, medium and large effect.
What to look for. Recommender is the large effect. Voice alone is nothing, and its interaction with the recommender is small but real.
| Effect | Compared | F (df) | p | Partial η² |
|---|---|---|---|---|
| Voice | between people | F(1, 725) = 0.46 | .497 | .001 |
| Recommender | within people | F(2, 1450) = 144.58 | < .001 | .166 |
| Recommender × voice | within people | F(2, 1450) = 15.89 | < .001 | .021 |
Computed on the 727 people who rated all three recommenders, with recommender within people and voice between people.
What to look for. The recommender effect and the interaction agree closely. The voice effect differs because it was tested on everyone, not only on people who rated all three.
| Effect | Here | In the dissertation |
|---|---|---|
| Recommender | F(2, 1450) = 144.58 | F(2, 1514) = 145.40, partial η² = .161 |
| Recommender × voice | F(2, 1450) = 15.89 | F(2, 1512) = 16.32, partial η² = .021 |
| Voice | F(1, 725) = 0.46 | F(1, 850) = 0.12, p = .727 |
The dissertation's degrees of freedom imply more people than the 727 who rated all three.
What to look for. The doctor was ranked first most often. The humanlike AI was most often ranked second.
| Recommender | Ranked first | Second | Third | Share first | Mean rank |
|---|---|---|---|---|---|
| Doctor | 452 | 121 | 148 | 62.7% | 1.58 |
| Humanlike AI | 122 | 442 | 157 | 16.9% | 2.05 |
| Standard AI | 147 | 158 | 416 | 20.4% | 2.37 |
721 people ranked the three on trust. Friedman test: χ²(2) = 230.24, p < .001. A lower mean rank is better.
Figure 7. First-place rankings
Table 6 as a figure. Each bar is the share of people who ranked that recommender first in trust. Chance would be one third.
Does a voice matter?
Half of the people heard the recommendation spoken as well as reading it. If a voice makes an AI seem more human, the humanlike AI should gain, and the standard AI may lose.
What to look for. Compare the written and voiced rows within each recommender. Differences are small.
| Measure, recommender and version | n | Mean | SD | SE | 95% CI | Median | Min | Max |
|---|---|---|---|---|---|---|---|---|
| Trust/belief: Doctor, written only | 349 | 1.20 | 1.28 | 0.07 | [1.07, 1.34] | 1.50 | −3.00 | 3.00 |
| Trust/belief: Doctor, written and voiced | 505 | 1.30 | 1.28 | 0.06 | [1.19, 1.41] | 1.50 | −3.00 | 3.00 |
| Trust/belief: Humanlike AI, written only | 315 | 0.99 | 1.45 | 0.08 | [0.83, 1.15] | 1.00 | −3.00 | 3.00 |
| Trust/belief: Humanlike AI, written and voiced | 474 | 1.24 | 1.37 | 0.06 | [1.11, 1.36] | 1.50 | −3.00 | 3.00 |
| Trust/belief: Standard AI, written only | 297 | 0.62 | 1.63 | 0.09 | [0.44, 0.81] | 1.00 | −3.00 | 3.00 |
| Trust/belief: Standard AI, written and voiced | 430 | 0.36 | 1.72 | 0.08 | [0.20, 0.52] | 0.50 | −3.00 | 3.00 |
| Trust: Doctor, written only | 349 | 1.11 | 1.43 | 0.08 | [0.96, 1.27] | 1.00 | −3.00 | 3.00 |
| Trust: Doctor, written and voiced | 505 | 1.23 | 1.42 | 0.06 | [1.11, 1.35] | 1.00 | −3.00 | 3.00 |
| Trust: Humanlike AI, written only | 315 | 0.88 | 1.59 | 0.09 | [0.70, 1.06] | 1.00 | −3.00 | 3.00 |
| Trust: Humanlike AI, written and voiced | 474 | 1.18 | 1.49 | 0.07 | [1.04, 1.31] | 1.00 | −3.00 | 3.00 |
| Trust: Standard AI, written only | 297 | 0.57 | 1.69 | 0.10 | [0.38, 0.76] | 1.00 | −3.00 | 3.00 |
| Trust: Standard AI, written and voiced | 430 | 0.32 | 1.78 | 0.09 | [0.15, 0.49] | 0.00 | −3.00 | 3.00 |
| Belief: Doctor, written only | 349 | 1.29 | 1.31 | 0.07 | [1.15, 1.43] | 1.00 | −3.00 | 3.00 |
| Belief: Doctor, written and voiced | 505 | 1.37 | 1.33 | 0.06 | [1.25, 1.48] | 2.00 | −3.00 | 3.00 |
| Belief: Humanlike AI, written only | 315 | 1.10 | 1.51 | 0.08 | [0.94, 1.27] | 1.00 | −3.00 | 3.00 |
| Belief: Humanlike AI, written and voiced | 474 | 1.30 | 1.41 | 0.06 | [1.17, 1.43] | 2.00 | −3.00 | 3.00 |
| Belief: Standard AI, written only | 297 | 0.68 | 1.68 | 0.10 | [0.49, 0.87] | 1.00 | −3.00 | 3.00 |
| Belief: Standard AI, written and voiced | 430 | 0.40 | 1.77 | 0.09 | [0.23, 0.57] | 1.00 | −3.00 | 3.00 |
Figure 8. Trust and belief, doctor, by voice
Each dot is one person's average trust and belief for the doctor, split by whether they heard a voice.
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 |
|---|---|---|---|
| Written only | 349 | 1.2 | 1.1 to 1.3 |
| Written and voiced | 505 | 1.3 | 1.2 to 1.4 |
Figure 9. Trust and belief, humanlike ai, by voice
Each dot is one person's average trust and belief for the humanlike ai, split by whether they heard a voice.
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 |
|---|---|---|---|
| Written only | 315 | 1.0 | 0.8 to 1.1 |
| Written and voiced | 474 | 1.2 | 1.1 to 1.4 |
Figure 10. Trust and belief, standard ai, by voice
Each dot is one person's average trust and belief for the standard ai, split by whether they heard a voice.
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 |
|---|---|---|---|
| Written only | 297 | 0.6 | 0.4 to 0.8 |
| Written and voiced | 430 | 0.4 | 0.2 to 0.5 |
What to look for. The d column. The voiced humanlike AI gains and the voiced standard AI loses, but the intervals are close to zero.
| Comparison | Voiced, Written only: mean (SD) | Difference [95% CI] | t (df) | p | d [95% CI] |
|---|---|---|---|---|---|
| Trust and belief: Doctor: voiced versus written | 1.30 (1.28), 1.20 (1.28) | 0.10 [−0.08, 0.27] | 1.08 (747.81) | .280 | 0.08 [−0.06, 0.21] |
| Trust and belief: Humanlike AI: voiced versus written | 1.24 (1.37), 0.99 (1.45) | 0.25 [0.04, 0.45] | 2.38 (646.74) | .017 | 0.18 [0.03, 0.32] |
| Trust and belief: Standard AI: voiced versus written | 0.36 (1.72), 0.62 (1.63) | −0.26 [−0.51, −0.02] | −2.10 (655.74) | .036 | −0.16 [−0.30, −0.01] |
| Trust: Doctor: voiced versus written | 1.23 (1.42), 1.11 (1.43) | 0.12 [−0.08, 0.31] | 1.16 (745.41) | .248 | 0.08 [−0.06, 0.22] |
| Trust: Humanlike AI: voiced versus written | 1.18 (1.49), 0.88 (1.59) | 0.30 [0.07, 0.52] | 2.62 (643.48) | .009 | 0.19 [0.05, 0.34] |
| Trust: Standard AI: voiced versus written | 0.32 (1.78), 0.57 (1.69) | −0.25 [−0.51, 0.01] | −1.92 (657.77) | .055 | −0.14 [−0.29, 0.00] |
| Belief: Doctor: voiced versus written | 1.37 (1.33), 1.29 (1.31) | 0.08 [−0.10, 0.26] | 0.84 (752.93) | .402 | 0.06 [−0.08, 0.19] |
| Belief: Humanlike AI: voiced versus written | 1.30 (1.41), 1.10 (1.51) | 0.20 [−0.01, 0.41] | 1.85 (639.7) | .065 | 0.14 [−0.01, 0.28] |
| Belief: Standard AI: voiced versus written | 0.40 (1.77), 0.68 (1.68) | −0.28 [−0.53, −0.02] | −2.14 (657.67) | .032 | −0.16 [−0.31, −0.01] |
| Doctor: mind perception, voiced versus written | 6.44 (2.22), 6.28 (2.28) | 0.15 [−0.16, 0.46] | 0.96 (721.87) | .339 | 0.07 [−0.07, 0.21] |
| Humanlike AI: mind perception, voiced versus written | 6.41 (2.35), 6.16 (2.43) | 0.25 [−0.09, 0.60] | 1.44 (644.78) | .150 | 0.11 [−0.04, 0.25] |
| Standard AI: mind perception, voiced versus written | 5.23 (2.68), 5.76 (2.66) | −0.53 [−0.93, −0.13] | −2.60 (620.54) | .010 | −0.20 [−0.35, −0.05] |
| All three recommenders averaged: voiced versus written | 0.98 (1.22), 0.92 (1.28) | 0.06 [−0.12, 0.25] | 0.67 (614.95) | .502 | 0.05 [−0.10, 0.20] |
Welch t tests, two-sided, uncorrected. d is positive when the voiced group is higher. A voice effect was not preregistered for each recommender separately.
Figure 11. Effect sizes of voice
Table 8 as a figure. Each dot is Cohen's d for voiced against written only, with a 95% interval.
What to look for. The same pattern appears with and without a voice: standard AI lowest, doctor highest.
| Comparison | n | First, second: mean (SD) | Difference [95% CI] | t (df) | p | Bonferroni p | dz [95% CI] | Higher / same / lower |
|---|---|---|---|---|---|---|---|---|
| Written only: Doctor versus standard AI | 297 | 1.16 (1.32), 0.62 (1.63) | 0.54 [0.37, 0.71] | 6.21 (296) | < .001 | 0.36 [0.24, 0.48] | 128 / 117 / 52 | |
| Written only: Humanlike versus standard AI | 297 | 0.97 (1.48), 0.62 (1.63) | 0.35 [0.20, 0.49] | 4.70 (296) | < .001 | 0.27 [0.16, 0.39] | 112 / 136 / 49 | |
| Written only: Humanlike AI versus doctor | 315 | 0.99 (1.45), 1.19 (1.30) | −0.20 [−0.31, −0.08] | −3.29 (314) | .001 | −0.19 [−0.30, −0.07] | 67 / 149 / 99 | |
| Written and voiced: Doctor versus standard AI | 430 | 1.34 (1.27), 0.36 (1.72) | 0.98 [0.82, 1.13] | 12.29 (429) | < .001 | 0.59 [0.49, 0.69] | 244 / 121 / 65 | |
| Written and voiced: Humanlike versus standard AI | 430 | 1.25 (1.36), 0.36 (1.72) | 0.89 [0.74, 1.03] | 12.27 (429) | < .001 | 0.59 [0.49, 0.69] | 229 / 156 / 45 | |
| Written and voiced: Humanlike AI versus doctor | 474 | 1.24 (1.37), 1.33 (1.26) | −0.09 [−0.18, −0.01] | −2.13 (473) | .034 | −0.10 [−0.19, −0.01] | 117 / 215 / 142 |
Paired t tests on trust and belief, uncorrected.
Figure 12. The same, by voice group
Table 9 as a figure. Each dot is dz with a 95% interval.
Mind perception
Mind perception is how much people credit a recommender with a mind: whether it can understand, feel, anticipate and explain itself. It is the mean of seven items, with Cronbach's alpha of .963 (2,304 ratings).
What to look for. The standard AI is lowest on every item. The doctor and humanlike AI are close.
| Item | Doctor | Humanlike AI | Standard AI |
|---|---|---|---|
| Capable of accurate recommendations | 6.54 | 6.37 | 5.50 |
| Can anticipate my health needs | 6.36 | 6.33 | 5.34 |
| Understands my feelings | 5.99 | 5.90 | 4.97 |
| Can account for my health needs | 6.49 | 6.41 | 5.50 |
| Process feels familiar | 6.17 | 6.26 | 5.43 |
| Process is transparent | 6.57 | 6.50 | 5.65 |
| I understand the process | 6.53 | 6.43 | 5.60 |
Mean rating of each item.
Figure 13. Each item, by recommender
Table 10 as a figure. Each dot is a mean with its 95% interval.
What to look for. Every correlation is positive. Mind perception goes with trust, warmth and competence.
| Pair | r |
|---|---|
| Trust/belief and Mind perception | .671 |
| Trust/belief and Warmth | .619 |
| Trust/belief and Competence | .668 |
| Mind perception and Warmth | .715 |
| Mind perception and Competence | .748 |
| Warmth and Competence | .782 |
Pearson correlations across all 2,298 ratings, pooled over recommenders.
Figure 14. Correlation grid
Table 11 as a figure. Darker means the two ratings move together more strongly.
Models
Mixed models predict each person's trust and belief rating, with a random intercept for each person. The reference is the humanlike AI read without a voice. Adding mind perception, warmth and competence shrinks the recommender coefficients.
What to look for. Read the standardized β column for which ratings carry the most weight.
| Predictor | b [95% CI] | Standardized β | z | p |
|---|---|---|---|---|
| Intercept | 1.001 [0.843, 1.158] | 12.45 | < .001 | |
| Standard AI (versus humanlike AI) | −0.349 [−0.500, −0.199] | −0.11 | −4.54 | < .001 |
| Voiced (versus written only) | 0.218 [0.015, 0.422] | 0.07 | 2.10 | .036 |
| Standard AI × voiced | −0.534 [−0.729, −0.338] | −5.35 | < .001 | |
| Doctor (versus humanlike AI) | 0.201 [0.055, 0.348] | 0.07 | 2.70 | .007 |
| Doctor × voiced | −0.122 [−0.311, 0.067] | −1.27 | .205 |
2,370 ratings from 854 people. Person variance 1.21, residual 0.89. Share of variance between people (ICC) .57. R² .06 from the predictors alone, .60 with the person effects.
Table 12 as a figure. Each dot is a coefficient (b) with its 95% interval.
What to look for. Read the standardized β column for which ratings carry the most weight.
| Predictor | b [95% CI] | Standardized β | z | p |
|---|---|---|---|---|
| Intercept | −1.484 [−1.666, −1.302] | −15.99 | < .001 | |
| Standard AI (versus humanlike AI) | −0.197 [−0.330, −0.063] | −0.06 | −2.88 | .004 |
| Voiced (versus written only) | 0.144 [−0.013, 0.301] | 0.05 | 1.80 | .073 |
| Standard AI × voiced | −0.215 [−0.390, −0.041] | −2.42 | .015 | |
| Doctor (versus humanlike AI) | 0.163 [0.034, 0.291] | 0.05 | 2.48 | .013 |
| Doctor × voiced | −0.116 [−0.283, 0.050] | −1.37 | .172 | |
| Mind perception | 0.401 [0.379, 0.423] | 0.66 | 35.82 | < .001 |
2,304 ratings from 852 people. Person variance 0.53, residual 0.68. Share of variance between people (ICC) .44. R² .47 from the predictors alone, .70 with the person effects.
Table 13 as a figure. Each dot is a coefficient (b) with its 95% interval.
What to look for. Read the standardized β column for which ratings carry the most weight.
| Predictor | b [95% CI] | Standardized β | z | p |
|---|---|---|---|---|
| Intercept | −1.989 [−2.215, −1.762] | −17.17 | < .001 | |
| Standard AI (versus humanlike AI) | −0.169 [−0.306, −0.032] | −0.05 | −2.42 | .016 |
| Voiced (versus written only) | 0.073 [−0.085, 0.232] | 0.02 | 0.91 | .364 |
| Standard AI × voiced | −0.068 [−0.247, 0.112] | −0.74 | .461 | |
| Doctor (versus humanlike AI) | 0.085 [−0.053, 0.222] | 0.03 | 1.21 | .227 |
| Doctor × voiced | −0.075 [−0.252, 0.102] | −0.83 | .406 | |
| Mind perception | 0.262 [0.230, 0.294] | 0.43 | 16.04 | < .001 |
| Warmth | 0.168 [0.111, 0.225] | 0.15 | 5.75 | < .001 |
| Competence | 0.294 [0.228, 0.361] | 0.23 | 8.72 | < .001 |
| Age group (1 to 4) | 0.207 [0.149, 0.264] | 0.14 | 7.07 | < .001 |
| Man (versus woman) | −0.048 [−0.174, 0.078] | −0.01 | −0.74 | .458 |
2,003 ratings from 688 people. Person variance 0.40, residual 0.65. Share of variance between people (ICC) .38. R² .55 from the predictors alone, .72 with the person effects.
Table 14 as a figure. Each dot is a coefficient (b) with its 95% interval.
Does mind perception explain the gaps?
For each pair of recommenders I asked how much of the difference in trust and belief runs through mind perception. The method compares each person's two ratings and uses 3,000 bootstrap resamples of people for the intervals.
What to look for. The a × b column and the share. For the doctor and the humanlike AI against the standard AI, more than half of the gap runs through mind perception.
| Comparison | People | a | b | a × b [95% CI] | Direct [95% CI] | Total | Share through mind | Dissertation a × b |
|---|---|---|---|---|---|---|---|---|
| Doctor versus standard AI | 688 | 0.98 | 0.47 | 0.46 [0.38, 0.56] | 0.35 [0.25, 0.45] | 0.81 | 57% | 0.49 |
| Humanlike AI versus standard AI | 683 | 0.85 | 0.47 | 0.40 [0.31, 0.49] | 0.28 [0.18, 0.37] | 0.68 | 59% | 0.41 |
| Humanlike AI versus doctor | 752 | −0.09 | 0.16 | −0.01 [−0.03, −0.00] | −0.12 [−0.18, −0.05] | −0.13 | 10% | −0.01 |
Difference-score mediation for repeated measures (Montoya and Hayes). The dissertation prints the humanlike-versus-doctor effect with the opposite sign to the direction of the comparison used here.
Figure 15. Indirect effects as a figure
Table 15 as a figure. Each dot is the part of the gap that runs through mind perception, in points of trust and belief, with a 95% interval.
Figure 16. Doctor versus standard AI
Arrows show how the comparison changes trust and belief directly and through mind perception. The numbers on the arrows are estimates.
Figure 17. Humanlike AI versus standard AI
Arrows show how the comparison changes trust and belief directly and through mind perception. The numbers on the arrows are estimates.
Figure 18. Humanlike AI versus doctor
Arrows show how the comparison changes trust and belief directly and through mind perception. The numbers on the arrows are estimates.