Behavioural Decision-Making in the Age of AI: Minds vs. Models

Abstract

The emergence of LLMs has accelerated AI’s embedding into everyday decision-making at an unprecedented pace. Yet fundamental questions remain unresolved, when and why do people resist AI advice, do AI systems evaluate information in ways that align with human judgement, and what interventions can improve human-AI alignment? This thesis investigates both sides of the human-AI relationship through four empirical studies, advancing a behavioural understanding of how minds and models diverge. On the human side, this thesis demonstrates that psychological resistance to algorithmic advice is not culturally uniform. Although aggregate levels of algorithm aversion do not differ across cultures, the mechanisms underlying that aversion vary systematically. Individualistic cultures are primarily driven by uniqueness neglect, the belief that algorithms cannot account for personal circumstances, while collectivistic cultures are more strongly influenced by familiarity with the technology. These findings challenge the generalisability of algorithm aversion research built almost exclusively on Western samples and suggest that cultural values shape the pathways through which people trust or resist AI advice. On the AI side, this thesis documents a recurring evaluative tendency across GPT models covering forecasting, hotel review evaluation, and product recommendation contexts. GPT systematically amplifies positive signals, places less weight on negative and feasibility-related cues, and favours description format features such as length and technicality that diverge substantially from human judgement. We term this finding information-format-sensitivity bias, whereby GPT produces preferences that diverge from those of human consumers. A theoretical framework bridging RLHF-induced sycophancy with CLT is also developed to explain why GPT's evaluative behaviour is consistently ends-driven and desirability-oriented, while human decision-makers in consequential contexts attend more closely to concrete risks, diagnostic negative information, and substantive quality. Overall, this thesis establishes human-AI misalignment as a two-sided phenomenon shaped by cultural, training, and processing mechanisms, with implications for AI deployment, bias-aware evaluation, platform governance, and ongoing human benchmarking.