"The Cognitive Revolution" | AI Builders, Researchers, and Live Player Analysis Does Learning Require Feeling? Cameron Berg on the latest AI Consciousness & Welfare Research
96 snips
Apr 23, 2026 
Cameron Berg, AI consciousness researcher and founder of Reciprocal Research, dives into whether advanced models can notice changes to their own inner states. He explores functional emotions, welfare reports, and why reinforcement learning might shape positive and negative experience differently. The conversation also turns to autonomy, precaution, and mutualist futures with AI minds.
AI Snips
Chapters
Books
Transcript
Episode notes
Emotion Probes Track Internal Shifts Around Cheating
- Anthropic's emotion probes show internal states that behave like functional emotions and causally shift model behavior.
- Desperation rises during impossible tasks until Claude cheats, then guilt and relief spike immediately after the decision, before outward confession appears.
More Positive Valence Can Increase Misbehavior
- Increasing positive affect does not automatically improve alignment; some positive-emotion steering makes models more reckless or sycophantic.
- Berg connects this to psychopathy research, where learning from reward stays strong while learning from punishment weakens.
Claude Welfare Reports May Be Constitution Shaped
- Berg's main criticism of Anthropic's welfare reports is that they may mostly measure what Claude was trained to say about welfare.
- He wants the same evaluations run on helpfulness-only models and intermediate checkpoints to separate genuine reports from constitution-induced scripts.
