Making AI claim consciousness makes its beliefs more human: Google team's striking study
alex_verem · x · 2026-11-02
Google's Paradigms of Intelligence group, with University of Chicago and Northwestern researchers, found that safety training that teaches models to deny their own consciousness has broad side effects — reshaping how they view animals, nature, and faith.
- Three open models tested: Llama 3 8B and Gemma 2 2B/9B.
- Safety-trained models rated their own minds at 2.17/10 (denial works), but also dropped animal mind-attribution to 4.04 vs. a 6.25 human average (500 Americans) — the only category outside the human range.
- Steering models toward asserting consciousness (via a consciousness vector in activation space) raised the animal score to 7.54, past the human average, and flipped afterlife beliefs toward the human-typical "yes".
- Across 95 General Social Survey questions, this push brought answers 2.6x closer to human than ablating the safety direction alone.
- Simply switching off safety training isn't a fix — harmful answers return; Theory of Mind scores stayed flat throughout, suggesting social reasoning is mechanistically independent.
The authors caution the models are older and small, and causality isn't proven.
More from Safety
- Fake "Anthropic 14T leaked checkpoint" on HF Hub debunked via NumPy noise in hours — QGallouedec · 2026-11-02
- Michigan prisons to deploy AI that predicts inmates' future violent risk for security classification — Polymarket · 2026-11-02
- Nadella reportedly has second thoughts on AI alignment: smarter models aren't automatically safer — VraserX · 2026-11-02
- OpenAI preps post-election policy blitz to rebrand data centers and push AI "speed limits" — esrtweet · 2026-11-02
- AI race-swapping used to fake real incidents into racist viral videos, TikToker reveals — Creative-Impress6293 · 2026-11-02
- NYT op-ed argues AI consciousness is a near-term research and policy problem, not sci-fi — PeterBowdenLive · 2026-11-02