Interacting with fallible AI: is distrust helpful when receiving AI misclassifications?

Due to the application of artificial intelligence (AI) in high-risk domains such as law and medicine, trustworthy AI and trust in AI are increasingly relevant to science and the public. A typical conception, for example, in the context of medical diagnosis, is that a knowledgeable user receives AI-generated classifications as advice. Research to improve such interactions often aims to foster users' trust, which, in turn, should improve combined human-AI performance. Given that AI models can err, we argue that the possibility of critically reviewing and thus distrusting an AI decision is an equally interesting target for research. We created two image classification scenarios in which participants received mock-up AI advice. The quality of the advice decreases during a phase of the experiment. We studied task performance, as well as participants' trust and distrust, and tested whether an instruction to remain skeptical and to review each piece of advice led to better performance compared to a neutral condition. Our results indicate that this instruction does not improve but rather worsens the participants' performance. Repeated single-item self-reports of trust and distrust indicate an increase in trust and a decrease in distrust following the decline in AI's classification quality, with no difference between the two instructions. Furthermore, through a Bayesian Signal Detection Theory analysis, we provide a procedure to assess appropriate reliance in detail by quantifying whether the issues of under- and over-reliance have been mitigated. We discuss the implications of our results for the usage of disclaimers before interacting with AI, as prominently used in current LLM-based chatbots, and for trust and distrust research.