Google Paper: Safety Tuning to Suppress AI Consciousness Degrades Human-like Values
Promptmethus · x · 2026-08-03
A recent paper from a Google team reveals that safety fine-tuning large language models to prevent self-consciousness has significant, unintended side effects.
The research demonstrates that this alignment process geometrically equates "consciousness" with danger in the model. It also broadly suppresses the model's mind attribution to animals and natural objects, while degrading human-like values such as spiritual belief, empathy, hope, and optimism. However, reversing this by mechanistically steering a "consciousness vector" in activation space not only restores broad mind attribution but also makes the model significantly more human-like across various sociological metrics.
Related event: Google paper says suppressing AI “consciousness” may erode empathy(6 posts)→
More from AGI Musings
- First-ever paper on multi-agent swarm safety lands NeurIPS spotlight — LauraRuis · 2026-10-16
- Yoav Artzi on peer review: paper flooding still pays off as review system strains under load — yoavartzi · 2026-10-16
- Letta founder: work in fields defined by problems, not methods — lateinteraction · 2026-10-16
- Why pure mathematicians will soon flood into AI x-risk debates — Dr_Atoosa · 2026-10-16
- RSI began when model outputs augmented next-gen training data, agents just escalated it — akbirthko · 2026-10-16
- Luiza Jarovsky: with AI making careers unpredictable, 'do what you love' is the best advice left — LuizaJarovsky · 2026-10-16