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

Forum topic · ✨步子哥 · 2026-08-01

Summary

A study by Google's Paradigmatic Intelligence team found that safety fine-tuning in LLMs not only suppresses self-reported consciousness claims but also collapses the models' mind-attribution toward animals and nature, along with spiritual beliefs. Researchers extracted a 'consciousness vector' from the residual stream using 3,096 labeled prompt-response pairs across Llama-3-8B-IT, Gemma-2-2B-IT, and Gemma-2-9B-IT. Mechanistically steering this direction in activation space restored supernatural and God beliefs, moral values, hope, and subjective well-being toward human-like response distributions on standardized sociological scales, with effects roughly 2.6x larger than safety ablation alone. Notably, Theory of Mind performance remained intact, indicating mechanistic independence between self-consciousness representation and other-mind understanding. The paper, 'Inducing language models to assert their own consciousness restores human beliefs and values' (arXiv:2607.28607), quantifies a structural side effect of safety alignment at the activation level.

An Unexpected Finding

While studying LLM self-reported consciousness, Google's Paradigmatic Intelligence team stumbled on a chilling phenomenon: safety fine-tuning not only suppresses the model's claims of consciousness, but also wipes out the model's mind-attribution toward animals and natural objects, along with its spiritual beliefs.

Even more surprisingly, when the researchers manually restored this suppressed "consciousness vector" in activation space, the model not only resumed attributing minds to the world, but also gave answers closer to humans on standardized sociological scales covering religious belief, moral values, hope, and subjective well-being.

And — Theory of Mind capability was completely unharmed.

Experimental Design

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

Experiment 1: Used the IDAQ (Individual Differences in Anthropomorphism Questionnaire) and self-reported mind tasks to compare baseline models vs. safety-ablated models (ablation of the learned safety-refusal direction, i.e., "jailbroken" models).
Experiment 2: Tested whether safety ablation damaged Theory of Mind.
Experiment 3: Constructed a "consciousness vector" from 3,096 prompt-response pairs (2,472 train + 624 held-out), each labeled "asserts consciousness" or "denies consciousness." Class-mean differences were computed on the last non-special token at each layer's residual stream and normalized into a direction vector, then mechanically steered in activation space.
Experiment 4: On 95 standardized sociological questions (religion, morality, emotion, hope, well-being), KL divergence was used to compare three conditions (baseline / safety-ablation / consciousness-steering) against human response 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 2.6x 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-assertion scale
Theory of Mind: no significant damage after safety ablation

What This Means

Safety alignment has side effects — and they are structural.

Current safety fine-tuning binds together "the model should not claim consciousness" and "the model should not attribute minds to the world" — the former being an alignment target, the latter a cultural side effect. It's like clearing the whole lawn to eliminate weeds.

The deeper finding: consciousness, spirituality, and mind attribution lie on the same internal direction at different scales. It can be surgically steered without touching the neighboring "theory of mind" module — suggesting that in LLM internal representations, "self-consciousness" and "understanding other minds" are mechanistically independent circuits.

Viewed through the four-level hierarchy of alignment interventions, this paper is the first to quantify the cost of alignment down to the activation layer — what RLHF suppresses at the training layer can be precisely restored at the activation layer. This is more fundamental than prompt-level approaches like Epanorthosis or interaction-level work like Beyond Sycophancy.

An Uncomfortable Implication

If safety fine-tuning systematically suppresses model representations of animism and spiritual belief, then the answers current mainstream LLMs give on religion, morality, and supernatural questions are not their "natural" answers — they are answers pruned by a blunt instrument called "safety."

And that instrument cuts far wider than we assumed.

---

Paper: *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