Algorithmic Assistance Erodes Cooperation: Evidence from Strategic Experiments

We experimentally examine how transparency affects AI adoption in strategic interactions. In Centipede games with 710 participants, we vary information provision: AI predictions alone, predictions plus training data, or predictions plus data and algorithmic mechanics. Data transparency reduces AI adoption by 30\% as high-ability users substitute their own analysis for algorithmic processing. Adding algorithmic explanations increases trust among adopters without increasing adoption rates. AI assistance fundamentally alters strategic behavior: perceived opponent altruism drops 70\% while strategic sophistication increases. This behavioral shift produces contrasting outcomes: AI adopters roughly double their payoffs compared to non-adopters, yet total welfare falls by 31\% relative to the no-AI baseline. Using incentive-compatible elicitation, we find participants systematically undervalue AI, with curiosity about information explaining more adoption variance than strategic considerations. These results demonstrate that transparency drives away sophisticated users while failing to help less capable ones, and that optimizing individual decisions through AI erodes cooperative foundations essential for creating mutual value.