Hanna Schraffenberger

Many decision support systems (DSS) provide predictions, recommendations, and, increasingly, explanations. Supporting human-AI decision-making with context-specific questions, however, remains largely unexplored. Questions can stimulate reflection and critical thinking, thereby introducing productive friction in the decision-making process and potentially reducing overreliance on DSS. This paper presents a proof-of-concept for generating data-driven questions based on a DSS prediction and its corresponding explanation, i.e., feature contribution, using a local language model. We illustrate our method using a realistic example from the medical field. In informal discussions (n = 2), gathering views on the possible usefulness of questions in decision-making, the clinicians mentioned that the generated questions have potential to help them reconsider the prediction and consider alternative options. Our proof-of-concept informs the design of human-AI interactions aimed at promoting the cognitive engagement of decision-makers and mitigating overreliance on DSS by shifting the focus from explanations to questions.

Superhuman AI that surpasses human performance in domains once regarded as uniquely human is increasingly infiltrating people’s daily lives. However, research suggests that people do not like being ‘outsmarted’ by AI. Although knowledge about the psychological impact of interacting with superhuman AI is crucial, empirical research is scarce. We conduct a 1-factorial between-subjects experiment in which participants compete against a purported AI opponent in a visual reasoning challenge. We measure the impact of being outperformed by AI on key psychological variables Sense of Agency and Self-Efficacy. Results show that being outperformed by AI leads to a lowered Sense of Agency and Task-Specific Efficacy. These, in turn, increase AI anxiety. This research sheds light on the psychological consequences of interacting with superhuman AI, and raises the question: if short exposure can already influence Sense of Agency and Task-Specific Efficacy, what does long-term and multiple exposure do to how humans perceive themselves?

The choices designers make about user interfaces can influence the choices end-users make when using those interfaces. These steering qualities of design have received much attention in the context of nudging and manipulative design (e.g., so-called ‘dark patterns’). This paper discusses ethical design practices and reviews the differences between nudging and manipulative design strategies, focusing on aspects that make them (un)acceptable. We show that existing distinctions are inconsistent and sometimes contradictory, often offering useful but partial or hard-to-follow advice. We illustrate these points with three concrete dilemmas that well-meaning designers face. Our review reveals that nudging and manipulative design both treat users as irrational and plagued by cognitive biases, which can be mitigated or exploited. We question this view and propose developing new interface design principles that treat users as rational and help them make their own deliberate autonomous decisions.

Social media platform’s lack of control over its content made way to the fundamental problem of misinformation. As users struggle with determining the truth, social media platforms should strive to empower users to make more accurate credibility judgements. A good starting point is a more accurate perception of the credibility of the message’s source. Two pre-registered online experiments (N = 525;N = 590) were conducted to investigate how verified source information affects perceptions of Tweets (study 1) and generic social media posts (study 2). In both studies, participants reviewed posts by an unknown author and rated source and message credibility, as well as likelihood of sharing. Posts varied by the information provided about the account holder: (1) none, (2) the popular method of verified source identity, or (3) verified credential of the account holder (e.g., employer, role), a novel approach. The credential was either relevant to the content of the post or not. Study 1 presented the credential as a badge, whereas study 2 included the credential as both a badge and a signature. During an initial intuitive response, the effects of these cues were generally unpredictable. Yet, after explanation how to interpret the different source cues, two prevalent reasoning errors surfaced. First, participants conflated source authenticity and message credibility. Second, messages from sources with a verified credential were perceived as more credible, regardless of whether this credential was context relevant (i.e., virtual lab coat effect). These reasoning errors are particularly concerning in the context of misinformation. In sum, credential verification as tested in this paper seems ineffective in empowering users to make more accurate credibility judgements. Yet, future research could investigate alternative implementations of this promising technology.

Documents are largely stored and shared digitally. Yet, digital documents are still commonly signed using (copies of) handwritten signatures, which are sensitive to fraud. Though secure, cryptography-based signature solutions exist, they are hardly used due to usability issues. This paper proposes to use digital identity wallets for securely and intuitively signing digital documents with verified personal data. Using expert feedback, we implemented this vision in an interactive prototype. The prototype was assessed in a moderated usability test (N = 15) and a subsequent unmoderated remote usability test (N = 99). While participants generally expressed satisfaction with the system, they also misunderstood how to interpret the signature information displayed by the prototype. Specifically, signed documents were also trusted when the document was signed with irrelevant personal data of the signer. We conclude that such unwarranted trust forms a threat to usable digital signatures and requires attention by the usable security community.