Artificial Intelligence in Human Domains — Experimental Evidence

This dissertation examines how individuals interact with artificial intelligence (AI) as it becomes embedded in human domains, such as moral decision-making. It combines preregistered experiments with an analysis of ChatGPT prompts to study (i) moral delegation, (ii) the demand for explanations for moral AI decisions, and (iii) needs for anthropomorphic design in large language model (LLM) conversational agents for human-like versus computer-like tasks.

The first chapter investigates delegation of a real, other-regarding donation decision in two experiments ($N = 5,639$). Participants either decide themselves or delegate to an AI or another human. Delegation demand is higher when the available delegate is AI, contrary to moral-domain algorithm aversion accounts. Delegation lowers perceived responsibility, more strongly so for AI delegates, and when stakes are real, participants adjust beliefs about AI capability in a self-serving manner that can rationalize offloading. These results indicate that AI can facilitate moral outsourcing and accountability diffusion, particularly when system opacity enables self-justification through belief adaptation about AI capabilities.
... mehr

The second chapter examines, in two experiments ($N = 393$; $N = 492$), when users access explanations of AI decisions in moral versus neutral contexts, and how affect and motivation shape this information choice. After observing an AI decision in either a trolley-type moral scenario or a matched neutral scenario, participants can choose to view an explanation. Average explanation uptake is high, but moral contexts increase anticipated psychological costs (e.g., anxiety and cognitive dissonance) and the motivation to protect existing attitudes. Curiosity and accuracy-oriented motivation are linked to higher explanation demand. Defense motivation is linked to lower demand, with particularly pronounced avoidance when the AI’s decision conflicts with a person’s own judgment, producing patterns suggestive of selective engagement.

The third chapter combines three controlled experiments ($N=624$) with an analysis of 5,394 ChatGPT user prompts to identify task-dependent anthropomorphism preferences. Users prefer more experience-related cues (e.g., warmth, empathy) for human-like tasks but less for computer-like tasks, while desired agency remains comparatively stable. Although ChatGPT adapts anthropomorphic style by task, this adaptation is systematically misaligned with user preferences. Human-like tasks increase mind perception yet reduce trust, highlighting the need for deliberately calibrated, task-sensitive conversational agent design.

Overall, the dissertation advances research on AI in human domains by showing that AI can be attractive for moral delegation and responsibility offloading, that governance through transparency depends on motivated user engagement, and that anthropomorphic cues should be matched to task-specific needs. These insights inform both AI governance (accountability, oversight, explainability) and the design of widely deployed conversational systems in contexts with ethically relevant real-world consequences.