Over the last two decades, researchers have demonstrated that affective states such as boredom, frustration, confusion, and engaged concentration can be detected with reasonable accuracy across diverse platforms, populations, and educational contexts. These detectors have enabled substantial advances in understanding how affect shapes learning. Yet no digital learning platform in use at scale today employs affect detection to drive adaptation or reporting. This talk examines that paradox. I will trace the development and spread of sensor-free affect detection in education and contrast its trajectory with knowledge tracing and at-risk prediction—two educational data mining technologies that have successfully reached large-scale deployment. I then present seven hypotheses for the gap: 1) insufficient performance for real-time decision-making, 2) poor generalizability, 3) high training costs, 4) limited perceived value, 5) engineering complexity, 6) the availability of universal alternatives, and 7) ethical and cultural resistance. These barriers are situated within the histories of other technologies that stalled or scaled slowly, from expert systems to jetpacks to voice recognition. I discuss candidate solutions and argue that the most promising near-term directions involve reducing per-deployment detector cost, designing fail-soft interventions, embedding affect detection within existing school practices, and leveraging generative AI to address multiple barriers simultaneously.