Situated within a collaborative research centre (CRC / TRR 318) concerned with Explainable AI, we study healthy distrust in machine learning results and explanations. Opposed to the typical concern for trust in AI, we focus on distrust in AI. Given that AI models can err, we propose that the possibility to critically review, to distrust, an AI decision and its explanation also needs to be considered.
Currently, we define healthy distrust as a temporary state of increased attention and vigilance towards a possible inadequate decision. We plan to study visual attention in the context of image classification. Participants will have to indicate their (dis)trust towards the classification explicitly. Furthermore, by following the Theory of Visual Attention (TVA; Bundesen, 1990) and by incorporating temporal-order judgment tasks in the presentation of the image classifications, attentional parameters will be assessed. To know and manipulate whether and to which degree errors are present in the classifications, the image (mis)classification will not be an output of an AI model but will be constructed. Estimates of the visual attention when correct or wrong classifications are presented will be assessed and compared to the participants’ explicit judgments of the classifications.