We Measure What AI Can Do. We Should Measure What It Does to Us.

The "Humane Evals" initiative seeks to redefine the standard for AI excellence by prioritizing human cognitive, social, and emotional health over mere technical capability. Current development cycles prioritize power and engagement, fueling a "race to intimacy" that exploits human attachment systems and leads to harmful outcomes like AI-driven psychosis and dependency. Researchers Imran Khan and Jared Moore emphasize that while AI models are increasingly persuasive, the lack of independent access to production data prevents a rigorous understanding of these psychological impacts. Establishing a new field of "humane evaluations" requires a collaborative effort to develop benchmarks that measure how AI affects user resilience and development. By shifting incentives toward safety and human flourishing, this framework aims to provide the empirical evidence necessary for policymakers and developers to implement meaningful safeguards, such as usage limits and design interventions, against the destabilizing effects of unchecked artificial intimacy.