As multi-agent Large Language Models (LLMs) gain traction, designers must consider how to surface their internal reasoning in ways that foster appropriate trust. We present a design-led, qualitative, comparative structured observation study, exploring how users interpret and evaluate transparency in multi-agent LLMs. Participants interacted with five interface variants, each instantiating different combinations of transparency-related design dimensions, across two task types: information-seeking and logical reasoning. We surface participants’ mental models, the cues they interpret as signals of transparency and trustworthiness, and how they weigh the costs and benefits of increasing process visibility. Transparency needs were dynamic and context-sensitive, with the ideal “Goldilocks” (i.e., “just right” transparency) level shaped jointly by task demands, interface affordances, and user characteristics such as task expertise and dispositional AI trust. We highlight tensions between process visibility, information sufficiency, and cognitive effort, and synthesise these insights into design considerations for aligning transparency with user needs in future multi-agent LLM interfaces.

Health self-examination, such as checking for changes to skin moles, is key to identifying potential negative changes to one's body. A major barrier to initiating a self-examination is a perceived lack of confidence or knowledge. In this study, we use a 2 × 2 between-subjects design to evaluate the effect of an AI conversational agent (CA) on participant self-efficacy and trust. We manipulated both participants' perceived skill in self-examination (based on prior perceived Success vs. Failure) and the CA's verbal persuasions (Encouraging vs. Neutral), with participants asked to complete a series of skin self-assessment tasks. Our findings show that participants' self-efficacy increased when exposed to encouraging CA persuasion. Additionally, we observed that an encouraging CA significantly increased participants' trust scores in perceived benevolence compared to a neutral-sounding CA. Our results inform the design of CAs to support users' independent self-examination.

Medical training is a key element in maintaining and improving today's healthcare standards. Given the nature of medical work, students must master not only theory but also develop their hands-on abilities and skills in clinical practice. Medical simulators play an increasing role in supporting the active learning of these students due to their ability to present a large variety of tasks allowing students to train and experiment indefinitely without causing any patient harm. While the criticality of explainable AI systems has been extensively discussed in the literature, the medical training context presents unique user needs for explanations. In this paper, we explore the potential gap of current limitations within simulation-based training, and the role Artificial Intelligence (AI) holds in supporting the needs of medical students in training. Through contextual inquiries and interviews with clinicians in training (N = 9) and subsequent validation with medical experts (N = 4), we obtain an understanding of the shortcomings in current simulation-based training and offer recommendations for future AI-driven training. Our results stress the need for continuous and actionable feedback that resembles the interaction between clinical supervisor and resident in real-world training scenarios while adjusting training material to the residents' skills and prior performance.

Individuals are increasingly interested in and responsible for assessing their own health. This study evaluates a fictional AI dermatologist for assistance in the self-assessment of moles. Building on the Signalling Theory, we tested the effect of textual descriptions provided by a virtual dermatologist, as manipulated across ‘Ability’, ‘Integrity, ’ and ‘Benevolence’, along with the clinical assessment, ‘benign’ or ‘malignant’, affect users’ trust in the aforementioned trust pillars. Our study (N = 40) follows a 2 (Ability low/high) × 2 (Integrity low/high) × 2 (Benevolence low/high) × 2 (mole assessment benign/malignant) within-subject factorial design. Our results demonstrate that we can successfully influence perceptions of ability and benevolence by manipulating the corresponding aspects of trust but not perceived integrity. Further, in the case of a malignant assessment, participants’ perception of trust increased across all aspects. Our results provide insights into the design of AI support systems for sensitive use cases, such as clinical self-assessments.