A new study evaluates the reliability of AI metrics used to assess user states in adaptive systems, revealing that only 31 out of 213 metrics are stable enough for real-time interpretation. The findings suggest a need for more rigorous validation in AI design to ensure responsible use of these technologies.