主讲人：魏立佳，武汉大学

标题：AI Transparency Paradox: When Medical AI Explanations Help and When They Harm

时间：2025年9月23日，下午3:30-5:00

线下地点：弘远楼311室

摘要：

We document a fundamental trade-off in algorithmic transparency: explanations for AI recommendations improve decision-making when algorithms are correct but systematically harm it when they err. Using a lab-in-the-field experiment with 257 medical students making 3,855 incentivized diagnostic decisions, we show that providing explanations increases diagnostic accuracy by 4.3 percentage points when AI advice is correct but decreases it by 4.6 percentage points when incorrect. This symmetric effect operates through indiscriminate increases in algorithmic reliance—explanations make all AI signals more persuasive regardless of quality. We develop a Bayesian framework showing participants treat explained AI as having 15.2 percentage points higher accuracy than its true rate of 73.4%, with this over-reliance persisting even for erroneous recommendations. The transparency paradox is most severe among uncertain decision-makers who need guidance most but are also most vulnerable to misleading explanations. Our welfare analysis reveals that contingent transparency policies—providing explanations only when AI confidence exceeds defined thresholds—generate 25-40% higher value than mandated universal transparency. These findings challenge the regulatory consensus that transparency universally improves human-algorithm collaboration and provide the first causal evidence that explanatory information from fallible sources can reduce welfare through asymmetric belief distortion.