The impact of explanations on layperson trust in artificial intelligence-driven symptom checker apps: experimental study

Abstract

Background: Artificial intelligence (AI)--driven symptom checkers are available to millions of users globally and are advocated as a tool to deliver health care more efficiently. To achieve the promoted benefits of a symptom checker, laypeople must trust and subsequently follow its instructions. In AI, explanations are seen as a tool to communicate the rationale behind black-box decisions to encourage trust and adoption. However, the effectiveness of the types of explanations used in AI-driven symptom checkers has not yet been studied. Explanations can follow many forms, including *why* -explanations and *how* -explanations. Social theories suggest that *why* -explanations are better at communicating knowledge and cultivating trust among laypeople. Objective: The aim of this study is to ascertain whether explanations provided by a symptom checker affect explanatory trust among laypeople and whether this trust is impacted by their existing knowledge of disease. Methods: A cross-sectional survey of 750 healthy participants was conducted. The participants were shown a video of a chatbot simulation that resulted in the diagnosis of either a migraine or temporal arteritis, chosen for their differing levels of epidemiological prevalence. These diagnoses were accompanied by one of four types of explanations. Each explanation type was selected either because of its current use in symptom checkers or because it was informed by theories of contrastive explanation. Exploratory factor analysis of participants' responses followed by comparison-of-means tests were used to evaluate group differences in trust. Results: Depending on the treatment group, two or three variables were generated, reflecting the prior knowledge and subsequent mental model that the participants held. When varying explanation type by disease, migraine was found to be nonsignificant (*P* =.65) and temporal arteritis, marginally significant (*P* =.09). Varying disease by explanation type resulted in statistical significance for input influence (*P* =.001), social proof (*P* =.049), and no explanation (*P* =.006), with counterfactual explanation (*P* =.053). The results suggest that trust in explanations is significantly affected by the disease being explained. When laypeople have existing knowledge of a disease, explanations have little impact on trust. Where the need for information is greater, different explanation types engender significantly different levels of trust. These results indicate that to be successful, symptom checkers need to tailor explanations to each user's specific question and discount the diseases that they may also be aware of. Conclusions: System builders developing explanations for symptom-checking apps should consider the recipient's knowledge of a disease and tailor explanations to each user's specific need. Effort should be placed on generating explanations that are personalized to each user of a symptom checker to fully discount the diseases that they may be aware of and to close their information gap. **J Med Internet Res 2021;23(11):e29386**