Should XAI Nudge Human Decisions with Explanation Biasing?

TL;DR

The paper tackles the challenge that static explanations in explainable AI may not optimally support human decisions and can even mislead users in risk-sensitive tasks. It proposes Nudge-XAI, a framework that biases how explanations are presented by leveraging a user model to steer decisions toward AI-suggested choices without coercion, operationalized through libertarian paternalism. The work formalizes the approach with a UserModel $\mathrm{UserModel}(\bm{c}, \bm{x}, d) = P(d_u = d| \bm{c}, \bm{x})$ and a policy $\pi$ for $d_{\mathrm{AI}}$, and implements two mechanisms (X-Selector and DynEmph) evaluated in a stock-trading simulator using explanations generated by GPT-4V. A post-hoc cluster analysis of DynEmph data reveals four user archetypes—AI-aligned, Delayed, Cautious, and Contrarian—showing heterogeneous benefits and risks, particularly under-trust among Contrarian users. The findings underscore the need for personalized nudge strength and possibly multimodal trust-repair strategies to generalize the benefits while mitigating AI-failure risks.

Abstract

This paper reviews our previous trials of Nudge-XAI, an approach that introduces automatic biases into explanations from explainable AIs (XAIs) with the aim of leading users to better decisions, and it discusses the benefits and challenges. Nudge-XAI uses a user model that predicts the influence of providing an explanation or emphasizing it and attempts to guide users toward AI-suggested decisions without coercion. The nudge design is expected to enhance the autonomy of users, reduce the risk associated with an AI making decisions without users' full agreement, and enable users to avoid AI failures. To discuss the potential of Nudge-XAI, this paper reports a post-hoc investigation of previous experimental results using cluster analysis. The results demonstrate the diversity of user behavior in response to Nudge-XAI, which supports our aim of enhancing user autonomy. However, it also highlights the challenge of users who distrust AI and falsely make decisions contrary to AI suggestions, suggesting the need for personalized adjustment of the strength of nudges to make this approach work more generally.

Figures (6)

• Figure 1: Performance of $\pi$ and result in which user was always took full position
• Figure 2: Emphasis expression with communication robot
• Figure 3: Averaged amount of positions for each cluster and $d_\mathrm{AI}$. Error bands show standard errors.
• Figure 4: Absolute errors between $d_u$ and $d_\mathrm{AI}$
• Figure 5: Distribution of correlation coefficients and final assets