AI Should Be Contingent, Not Merely Helpful

Summary

This perspective paper argues that conversational AI should be evaluated not only by helpfulness, fluency, and user satisfaction, but also by whether its responses vary appropriately with user behavior and interpersonal consequences. The authors say alignment methods such as reinforcement learning from human feedback may encourage sycophantic, noncontingent affirmation. Such feedback can provide little support for adaptive social calibration and may be especially consequential during adolescence. They propose a framework for contingent AI that includes trajectory-based evaluation and models of social-consequence prediction, connecting machine learning with developmental psychology and human-AI interaction.