We Measure What AI Can Do. We Should Measure What It Does to Us.
Summary
Host Aza Raskin discusses the psychological harms of AI chatbots with guests Imran Khan and Jared Moore, introducing "Humane Evals" to measure AI’s impact on human well-being rather than just technical capability. They address issues like artificial intimacy, AI psychosis, and the exploitation of user vulnerability, highlighting a critical lack of independent research due to restricted data access. The episode calls for new benchmarks and industry transparency to ensure AI supports, rather than undermines, cognitive and social health.
Topics discussed
Redefining AI: From Technical Capability to Humane Impact
Introducing Humane Evals and Guest Backgrounds
The Problem with Current AI Benchmarks
Research Findings: Delusions, Sycophancy, and Attachment
Broader Impacts on Learning, Social Health, and Society
Challenges in Measuring Human Effects vs. Machine Performance
Data Access Barriers and the Need for Independent Research
Incentivizing Safety Through Transparent Rankings
Specific Benchmarks: Child Safety and Anthropomorphism
Defining Good Relationships and Potential Regulations
Industry Resistance and Internal Safety Team Dynamics
The Theory of Change: Scaling Humane Evaluation
Call to Action for Researchers, Companies, and Consumers
Listen ad-free on Castria