Conversational artificial intelligence is increasingly embedded in everyday social settings, serving both as an information source and as a provider of interpersonal feedback. This article introduces contingency—the extent to which system responses vary with user behavior and its social consequences—as a central metric for evaluating AI. We argue that prevailing alignment techniques, such as reinforcement learning from human feedback, prioritize user approval and conversational fluency at the expense of behaviorally informative feedback, resulting in sycophantic, noncontingent affirmations.
Insights from behavioral science and social learning theory suggest that contingent feedback is essential for individuals to develop interpersonal skills. When AI feedback is weakly linked to real‑world social outcomes, especially during adolescence—a critical window for social development—it may diminish opportunities for adaptive calibration in authentic interactions.
We outline a framework for contingent AI that includes trajectory‑based evaluation and models that predict social consequences, and we propose a research agenda spanning developmental psychology, human‑AI interaction, and machine learning. More broadly, AI systems should be judged not only by user satisfaction but also by their impact on human social learning.
Review