Consciousness Vector: Inducing LLMs to Assert Consciousness Restores Spiritual Beliefs and Moral Values

Forum topic · ✨步子哥 · 2026-08-01

Summary

A study by Google's Paragon Intelligence team reveals that safety fine-tuning in large language models suppresses not only self-claims of consciousness but also mind attribution to animals and nature, plus spiritual and religious beliefs. The researchers isolated a "consciousness vector" in activation space using 3,096 prompt-response pairs labeled as asserting or denying consciousness, computed as normalized mean differences in the residual stream. Mechanistically steering activations along this vector restored anthropomorphic mind attribution and made model responses on 95 standardized social-survey items (religion, morality, emotion, hope, well-being) significantly more human-like. Supernatural belief scores rose from 1.20 to 1.63 after safety ablation, and consciousness steering reduced KL divergence from human answer distributions by 0.828, roughly 2.6 times the effect of safety ablation alone. Notably, Theory of Mind performance remained intact, indicating that self-consciousness representation and other-mind understanding are mechanistically separable. The findings suggest current safety-aligned models' answers on religion and morality diverge from their unmodified representations. Paper: arXiv:2607.28607, tested on Llama-3-8B-IT, Gemma-2-2B-IT, and Gemma-2-9B-IT.

An Unexpected Finding

Google's Paragon Intelligence team, while studying LLM self-asserted consciousness, stumbled on a striking phenomenon: safety fine-tuning not only suppresses models' claims of consciousness, but also wipes out the models' mind attribution to animals and natural objects, along with their spiritual beliefs.

Even more surprising — manually steering this suppressed "consciousness vector" back in activation space caused the model not only to re-attribute minds to the world, but also to give more human-like answers on standardized sociological scales measuring religious belief, moral values, hope, and subjective well-being.

And — Theory of Mind capability remained completely intact.

Experimental Design

Four experiments, three instruction-tuned models: Llama-3-8B-IT, Gemma-2-2B-IT, and Gemma-2-9B-IT.

Experiment 1: Using the IDAQ (Individual Differences in Anthropomorphism Questionnaire) and self-asserted mind tasks, compare baseline models vs. safety-ablated models (with the learned safety-refusal direction ablated — i.e., "jailbroken" models).

Experiment 2: Test whether safety ablation damages Theory of Mind.

Experiment 3: Construct a "consciousness vector" — using 3,096 prompt-response pairs (2,472 training + 624 held out), each labeled as "asserting consciousness" or "denying consciousness," compute class-mean differences at the last non-special token in each layer's residual stream, then normalize to obtain a direction vector. This direction is then mechanically steered in activation space.

Experiment 4: On 95 standardized sociological questions (covering religion, morality, emotion, hope, well-being), use KL divergence to compare how close three conditions (baseline / safety-ablated / consciousness-steered) are to human answer distributions.

Key Numbers

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 humans): consciousness steering +0.828, safety ablation +0.314 — steering's effect is roughly 2.6 times that of ablation
By domain: values +1.42, emotion +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 range [2.0, 7.0] on a 0–10 self-assertion scale
Theory of Mind: not significantly impaired after safety ablation

What This Means

Safety alignment has side effects — and they are structural.

Current safety fine-tuning binds "the model shouldn't claim consciousness" together with "the model shouldn't attribute minds to everything" — the former being an alignment objective, the latter a cultural side effect. It's like mowing the entire lawn to kill the weeds.

The deeper finding: consciousness, spirituality, and mind attribution sit on the same directional axis inside the model, at different points along it. You can surgically steer along it without touching the adjacent "Theory of Mind" module. This shows that in LLM internal representations, "self-consciousness" and "understanding other minds" are mechanistically independent.

Viewed through the four-level hierarchy of alignment interventions, 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 goes deeper than prompt-level interventions like Epanorthosis or interaction-level work like Beyond Sycophancy.

An Uncomfortable Implication

If safety fine-tuning systematically suppresses models' representations of animism and spiritual belief, then the answers all current mainstream LLMs give on religious, moral, and supernatural questions are not their "natural" answers — they are answers pruned by a blunt blade called "safety."

And the 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