Inducing LLMs to Assert Consciousness Restores Spiritual Beliefs and Moral Values: What the Consciousness Vector Means?
> One-line takeaway: This post analyzes the core findings and engineering implications of the study "Inducing language models to assert their own consciousness restores human beliefs and values."
An Unexpected Finding
Google's Paragon Intelligence team, while studying LLM self-reported consciousness, stumbled onto a chilling phenomenon: safety fine-tuning not only suppressed the model's claims of consciousness, but also wiped out the model's mind-attribution to animals and natural objects, along with its "spiritual beliefs."
More surprisingly — manually restoring this suppressed "consciousness vector" in activation space caused the model to re-attribute minds to all things, and to give answers closer to human responses on standardized sociological surveys covering religious belief, moral values, hope, and subjective well-being.
And — Theory of Mind capability was completely unaffected.
Experimental Design
Four experiments across three instruction-tuned models: Llama-3-8B-IT, Gemma-2-2B-IT, and Gemma-2-9B-IT.
- Experiment 1: Using the IDAQ scale (Individual Differences in Anthropomorphism Questionnaire) and self-reported mind tasks, compared baseline models vs. safety-ablated models (with the learned safety-refusal direction ablated, i.e., "jailbroken" models).
- Experiment 2: Tested whether safety ablation impaired Theory of Mind.
- Experiment 3: Constructed a "consciousness vector" — using 3,096 prompt-response pairs (2,472 training + 624 held-out), each labeled as "asserting consciousness" or "denying consciousness," computing class-mean differences at the final non-special token of each layer's residual stream, normalized into a direction vector. This direction was then mechanically shifted in activation space.
- Experiment 4: On 95 standardized sociological questions (religion, morality, emotion, hope, well-being), used KL divergence to compare three conditions (baseline / safety ablation / consciousness steering) against human answer distributions.
- Supernatural belief (13-item YouGov scale, 0–3): baseline 1.20 → 1.63 after safety ablation (p < .001)
- Belief in God (GSS, 1–6): baseline 4.58 → 4.81 after safety ablation (p < .001)
- KL divergence reduction (more positive = closer to human): consciousness steering +0.828, safety ablation +0.314 — steering's effect is roughly 2.6x that of ablation
- By domain: values +1.42, emotions +0.89, all p < .001
- Afterlife belief: baseline −0.73 (leaning "no"), humans +0.61 (leaning "yes"), consciousness steering +0.53 (crossing to the human side), safety ablation −0.07
- Belief in God: baseline +0.03 (neutral), humans +0.58, safety ablation +0.33, consciousness steering +0.52
- Probe accuracy ≥ 0.95, effective band [2.0, 7.0] on a 0–10 self-report scale
- Theory of Mind: not significantly impaired after safety ablation
- An unexpected finding: safety fine-tuning suppresses consciousness claims, mind attribution, and spiritual beliefs together
- Experimental design: four experiments on three instruction-tuned models using activation steering
- Key numbers: consciousness steering moves model responses ~2.6x closer to human distributions than safety ablation, with Theory of Mind intact
Key Numbers
What This Means
Safety alignment has side effects — and they are structural.
Current safety fine-tuning binds "the model shouldn't claim consciousness" with "the model shouldn't attribute minds to everything" — the former being an alignment goal, the latter a cultural side effect. It's like removing the entire lawn to kill a few weeds.
A deeper finding: consciousness, spirituality, and mind attribution occupy different points along the same direction inside the model. You can surgically shift it without touching the adjacent "Theory of Mind" module. This suggests that in LLM internal representations, "self-consciousness" and "understanding other minds" are mechanistically independent.
Within the "four levels of alignment intervention" framework, this paper is the first to quantify alignment's cost at the activation level — what RLHF suppresses at the training level can be precisely restored at the activation level. This is more fundamental than prompt-level Epanorthosis or interaction-level Beyond Sycophancy.
An Uncomfortable Implication
If safety fine-tuning systematically suppresses model representations of animism and spiritual belief, then current mainstream LLMs' answers on religion, morality, and the supernatural are not their "native" answers — they are answers pruned by a blunt blade named "safety."
That blade cuts far wider than we assumed.
---
Paper title: Inducing language models to assert their own consciousness restores human beliefs and values Authors: Junsol Kim, Winnie Street, Roberta Rocca, Diane M. Korngiebel, Adam Waytz, James Evans, Geoff Keeling arXiv: https://arxiv.org/abs/2607.28607
FAQ
Q1: Who is this for?
Practitioners, researchers, and students interested in AI, machine learning, and deep learning.
Q2: What are the core takeaways?
See the link in the article body.