1 Introduction

Reference Longoni, Bonezzi and MorewedgeLongoni, Bonezzi and Morewedge (2019) recently reported a series of ten clever experiments demonstrating that people prefer to receive medical care (i.e., diagnosis, screening, and treatment) from a human provider rather than from an equally competent artificially intelligent computer (AI provider). Particularly interesting is their finding that perceived “uniqueness neglect” on the part of an AI provider can explain this preference.

There is an important message, however, that we believe became lost among the other interesting findings: People actually did prefer the AI provider so long as it outperformed the human provider. The effects of accuracy could not be tested in seven of their ten studies, either because participants received incomplete accuracy information that did not allow for a direct comparison between human and computer (Studies 1 and 4) or were explicitly told that the human and computer had equal accuracy rates (Studies 2, 5, 6, 7 and 8). We believe, however, that the findings of Studies 3A to 3C provide strong support that people do sometimes embrace AI.

Studies 3A to 3C were identical in design, differing only in the type of service provided (screening for skin cancer, triaging a potential emergency, surgical implant of pacemaker). Participants rated provider X (human) and Y (human or comupter) on a 7-point scale such that scores above the midpoint of the scale indicated a preference for Y, below the midpoint indicated a preference for X, and scores exactly at 4 indicated indifference. One variable that they manipulated was the relative accuracyFootnote ¹ rate of the X and Y providers. Accuracy rates were either equal (X = Y), slightly favored provider Y (X < Y), or strongly favored provider Y (X << Y). They also manipulated choice set: Provider X was always human, but provider Y was either another human or a computer (AI provider). Longoni et al. report significant main effects of both relative accuracy and choice set, but no interaction (see Figure 1). Before discussing the results, here is their stated goal for these three studies:

Figure 1: Preference for human vs. AI providers as a function of relative accuracy (adapted from Reference Longoni, Bonezzi and MorewedgeLongoni et al. 2019, Figure 2). A response of 4 indicates indifference.

First, their reported main effect shows that, overall, increases in accuracy for provider Y did lead to a preference for this provider. On the far left of each graph in Figure 1, we see that when accuracy rates are identical (X = Y), ratings are at or below the midpoint, indicating either indifference between humans (black bars) or a preference for the human over AI (white bars). This replicates the findings of studies 2, and 5–8. However, when accuracy rates are better for provider Y (X < Y or X << Y), mean responses increased above the midpoint of the scale indicating a preference for provider Y. The lack of an interaction showed this to be the case regardless of whether provider Y was another human (black bars) or a computer AI (white bars). Clearly, participants did not always resist AI technology. In fact, they arguably embraced AI when its historical performance was described as being better than that of the human (Reference Carlson, Desai, Drury, Kwak and YancoCarlson et al., 2011). Note that the effect size for accuracy is quite large, with partial eta squared values of η²_p = .445 to .479. These effects are 2 to 3 times larger than effects for choice set (η²_p = .131 to .222).

Using the means and standard deviations provided in their Table 2, and assuming that the sample size was equal across conditions, we calculated the one-sided lower 95% confidence limit on each reported mean. In all six conditions involving a superior automated provider, the lower bound for the mean response was above the null value of 4.0 (i.e., the value indicating no preference) suggesting that participants preferred the more accurate automated provider to the less accurate human one.