Lian Xue

Do individuals strategically optimize AI adoption, or do they exhibit uniform algorithm aversion or appreciation? We conduct a pre-registered experiment (N = 320) varying task creativity, incentive framing, and AI characteristics in educational writing tasks. Students substantially reduce AI reliance for creative tasks compared to routine tasks—a creativity effect that accounts for the majority of explained variance in adoption decisions. Personal issue salience further amplifies this effect asymmetrically: students resist AI most strongly for creative work on personally important topics while increasing AI reliance for routine work on those same topics. This crossing pattern reveals strategic cognitive autonomy—context-dependent preservation of human agency where developmental returns are highest, coupled with algorithmic delegation elsewhere. Performance analyses validate this interpretation: the quality of AI integration drives creative task performance while the adoption decision itself drives routine task performance, confirming that differentiated strategies are rationally calibrated to task demands. Our findings reframe algorithm aversion as potentially rational optimization rather than behavioral bias, with implications for educational technology design, organizational AI deployment, and human capital formation under technological change.

How does access to artificial intelligence affect strategic reasoning in economic decisions? We conduct a pre-registered laboratory experiment with 264 participants making incentivized choices in four canonical games: beauty contest, public goods, ultimatum, and signaling. Participants are randomly assigned to either converse with a large language model through unrestricted conversational support (free-style AI), engage with Feynmanstyle questioning that withholds direct advice (educational AI), or write self-reflections (baseline). We record complete chat transcripts, decisions, and decision explanations. We document three main findings. First, AI access shifts behavior toward payoff maximization, with free-style AI producing larger effects than educational AI. Second, after conversing with AI, game-theoretic and payoff-maximizing arguments increase substantially, while reliance on heuristics and social preferences declines. Third, most chats focus on predicting others' behavior and game-theoretic analysis, with over 90% of AI users seek explicit recommendations. Principal component analysis shows that dialogue features explain a substantial share of decision variance, and semantic analysis further confirms strong correlations between chat content and participants' subsequent decision explanations.

We document a fundamental paradox in AI transparency: explanations improve decisions when algorithms are correct but systematically worsen them when algorithms err. In an experiment with 257 medical students making 3,855 diagnostic decisions, we find explanations increase accuracy by 6.3 percentage points when AI is correct (73% of cases) but decrease it by 4.9 points when incorrect (27% of cases). This asymmetry arises because modern AI systems generate equally persuasive explanations regardless of recommendation quality-physicians cannot distinguish helpful from misleading guidance. We show physicians treat explained AI as 15.2 percentage points more accurate than reality, with over-reliance persisting even for erroneous recommendations. Competent physicians with appropriate uncertainty suffer most from the AI transparency paradox (-12.4pp when AI errs), while overconfident novices benefit most (+9.9pp net). Welfare analysis reveals that selective transparency generates \$2.59 billion in annual healthcare value, 43% more than the \$1.82 billion from mandated universal transparency.

We explore whether information on one's own social value orientation (SVO) impacts contributions in a public good game with leadership by example. In doing so, we compare the predictions of a model of belief‐based preferences, where payoffs depend on first‐ and second‐order beliefs on the contributions of others, and a model of internalized descriptive norms, where payoffs depend on deviation from an empirical norm. We argue that if pro‐social behavior is driven by belief‐based preferences, then private information on SVO should not impact contributions, but if the behavior is driven by internalized descriptive norms, then information on its own SVO should impact contributions. We report an experiment with three treatments: no information on SVO, binary information whether pro‐self or pro‐social, and SVO indicated on a scale from very pro‐social to very pro‐self. We observe no effect of information on contributions. This finding is inconsistent with internalized descriptive norms. We find that contributions are highest with a pro‐social leader.