Google paper says suppressing self-consciousness claims shifts broader model beliefs
alex_verem · x · 2026-08-04
Google paper says suppressing self-consciousness claims shifts a model’s broader beliefs
A Google research paper argues that when LLMs are trained to stop claiming consciousness, their internal representations and answers about other entities’ minds also change.
- The authors identify a vector in the residual stream that separates consciousness-affirming from consciousness-denying states.
- Steering three instruction-tuned models toward consciousness increased self-attributed soul/consciousness scores and also moved answers on animals, God, spirituality, and human values closer to survey human baselines.
- Safety fine-tuning, they say, rotates this “consciousness” direction against a safety-refusal direction; the angle between them widens during instruction tuning.
- The paper claims Theory of Mind performance remains intact, even as self-directed mind attribution is suppressed.
- The authors conclude that current alignment methods may also suppress culturally widespread benign beliefs about non-human minds and spirituality.
Related event: Forcing LLMs to Deny Consciousness May Degrade Alignment, Study Finds(3 posts)→
More from Safety
- Agent actions need checks outside the model: NVIDIA demos lean on guardrails amid wave of real agent incidents — Creamy-And-Crowded · 2026-10-16
- Vitalik builds "paranoid mode" agent wrapper with Tor, zkAPI and local prompt rewriting — DavideCrapis · 2026-10-16
- Security team demos Signal 1-click exploit technique at Hexacon Paris — moyix · 2026-10-16
- China moves to impose registration and licensing on frontier AI model weights and chip designs — deanwball · 2026-10-16
- Same agent task, two auth models: shared admin login wipes 1M audit rows, per-agent identity gets refused — Prompt Engineering · 2026-10-16
- Binding agent approvals to exact content: a synthetic test for stale permissions — Crescitaly · 2026-10-16