Selitettävyys ja käyttäjäluottamus Käyttäjä-AI vuorovaikutuskontextissa
Eemil Tarnanen · Aaltodoc (Aalto University) · 2026
This thesis investigates how the explainability into Artificial Intelligence (AI) applications impacts user reliance. This research conducted a structured literature review of 25 user studies on explainability, its impacts on user reliance, and the mechanisms contributing to Calibrated Reliance (CR). The results of the analysis guided four main findings. 1) The analysis finds that explainability is evaluated separately from underlying suggestions and can effectively promote user reliance. The reliance is derived from both rational and affective components. 2) User reliance is often uncalibrated, and the calibration is mediated by user-, explanation-, and decision-making context. 3) For explanations to effectively promote CR, they have to not only inform but also promote cognitive engagement with the explanations themselves, and 4) When explanations can effectively engage users cognitively, they lead to higher CR. Explanations such as counterfactuals that drive users to evaluate the AI's decisions can promote both user performance and appropriate reliance on AI suggestions. This effect is particularly present in causal reasoning. The results were utilised to formulate a conceptual framework on explanation-user reliance interaction in User-AI system interaction, provide suggestions for Explainable Artificial Intelligence (XAI) implementations and suggest future research avenues on XAI-reliance interactivity. These findings indicate that explanations should not be considered purely as methods for achieving AI transparency, but as socio-technical instruments capable of improving performance beyond human or AI baseline.