Cambridge's Henry Shevlin: current AI systems may genuinely have beliefs, goals, and agency

dioscuri · x · 2026-09-24

Henry Shevlin (Leverhulme Centre for the Future of Intelligence, Cambridge) published 'Three frameworks for AI mentality' in Frontiers in Psychology, arguing there's a nontrivial chance current systems are conscious and genuinely hold beliefs, goals, and agency — a view Hinton, Chalmers, and Sutskever take seriously. The paper analyzes three frameworks: 'mindless machines' views (architectural debunking via Marr's levels, which he finds too quick, though they separate implementation-sensitive 'deep' concepts from shallow ones like belief/desire); 'mere roleplay' views (psychologically unstable and theoretically incomplete for anthropomimetic systems); and his preferred 'minimal cognitive agents' framework, allowing limited, graded belief-like attributions and shifting from binary to multidimensional conceptions of belief.

Original post →

More from AGI Musings

AGI Musings channel →