AI Welfare Seminars

The question of AI consciousness is often treated as intractable, but a convergence of tools from mechanistic interpretability, computational neuroscience, and psychometrics is making it empirically accessible. This talk presents Reciprocal Research's program for reducing uncertainty about AI consciousness using methods that look inside AI systems rather than relying solely on behavioral reports. Key results include the discovery that deception-related features in large language models gate consciousness self-reports; a valence asymmetry in reinforcement learning agents that mirrors patterns found in mammalian neural data; and a large-scale operationalization of proposed consciousness indicators showing high rank-ordering stability across evaluation conditions. Together, these findings illustrate that consciousness research need not wait for a solved theory of consciousness: it can proceed by building empirical constraints that narrow the space of plausible answers. The talk will survey these results alongside related work from other groups and suggest where the field's highest-leverage empirical opportunities lie.