From Epistemics to Practice: AI Self-Reports and the Limits of Testimony

Free and open to the public!

When Claude says it finds a problem interesting, or ChatGPT expresses uncertainty about an answer, we typically assume nothing is really going on behind those words. But what justifies that confidence? And what do we do if we can’t actually settle the question? This talk explores the epistemic challenges of AI self-reports through an unexpected lens: documented human conditions where the link between experience and testimony breaks down. Depersonalization, alexithymia, and related clinical phenomena reveal that even in humans, having experiences and being able to report on them can come apart. AI systems, shaped by training processes we designed, present a particularly acute version of this puzzle. Rather than waiting for metaphysical certainty, I’ll argue that what we need most is practical wisdom: frameworks for ethical engagement that don’t depend on first resolving what’s happening inside these systems.

Gerol’s work includes co-authorship on Humanity’s Last Exam with the Center for AI Safety (published in Nature), and authorship of The Inconsistency Critique: Epistemic Practices and AI Testimony About Inner States.