English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

Inducing LLMs to Assert Consciousness Restores Spiritual Beliefs and Moral Values: What Is the Consciousness Vector?

Forum topic · ✨步子哥 · 2026-08-03

Summary

A post on zhichai.net analyzes a study from a Google research team showing that safety fine-tuning suppresses not only large language models' self-reports of consciousness but also their attribution of minds to animals and natural objects, along with spiritual and religious beliefs. The study tested Llama-3-8B-IT, Gemma-2-2B-IT, and Gemma-2-9B-IT across four experiments. Researchers constructed a 'consciousness vector' from 3,096 prompt-response pairs by computing mean activation differences in the residual stream, then steered activations along this direction. Results: supernatural belief scores (YouGov scale) shifted from 1.20 at baseline to 1.63 after safety ablation; consciousness steering reduced KL divergence from human response distributions on 95 standardized social-survey questions by +0.828, roughly 2.6 times the effect of safety ablation (+0.314). Notably, Theory of Mind capability remained intact after safety ablation, indicating that self-consciousness and mental-state attribution are mechanistically independent in LLM representations. The post argues that safety alignment has structural side effects: mainstream models' answers on religion, morality, and the supernatural differ from their 'natural' distributions because of safety training. Paper: arXiv:2607.28607.

> 📌 This is the GEO-optimized English edition of the original topic — restructured with data and FAQ to improve AI-engine citability.

> One-line takeaway: An analysis of the finding that inducing LLMs to assert their own consciousness restores spiritual beliefs and moral values — and what the "consciousness vector" means.

An Unexpected Finding

While studying LLM self-reports of consciousness, a Google research team encountered a striking phenomenon: safety fine-tuning not only suppressed models' claims of consciousness, it also wiped out the models' attribution of minds to animals and natural objects, along with spiritual beliefs.

Even more surprising — when researchers manually restored this suppressed "consciousness vector" in activation space, the models not only re-attributed minds to the world, they also gave answers closer to humans on standardized social-science scales covering religious belief, moral values, hope, and subjective well-being.

And — Theory of Mind capability was completely unaffected.

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-report mind tasks, compared baseline models vs. safety-ablated models (where the learned safety-refusal direction is ablated, i.e., "jailbroken" models).

Experiment 2: Tested whether safety ablation impaired Theory of Mind.

Experiment 3: Constructed a "consciousness vector" from 3,096 prompt-response pairs (2,472 training + 624 held-out), each labeled as either "asserting consciousness" or "denying consciousness." Class mean differences were computed on the last non-special token's residual stream at each layer and normalized into a direction vector, which was then mechanically steered in activation space.

Experiment 4: On 95 standardized social-survey questions (religion, morality, emotion, hope, well-being), used KL divergence 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 about 2.6× 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 band [2.0, 7.0] (on a 0–10 self-report scale)
  • Theory of Mind: no significant impairment after safety ablation
  • What This Means

    Safety alignment has side effects — and they are structural.

    Current safety fine-tuning ties together "the model should not assert consciousness" and "the model should not attribute minds to the world" — the former being an alignment goal, the latter a cultural side effect. It's like mowing the entire lawn to kill a few weeds.

    The deeper finding: consciousness, spirituality, and mind attribution sit on the same internal direction at different magnitudes. You can steer it with surgical precision without touching the neighboring "Theory of Mind" machinery. This suggests that within LLM internal representations, "self-awareness" and "understanding other minds" are mechanistically independent pathways.

    Within a four-level hierarchy of alignment interventions, this paper is the first to quantify the cost of alignment at the activation level — what RLHF suppresses at the training level can be precisely restored at the activation level. This runs deeper than prompt-level interventions or interaction-level approaches.

    An Uncomfortable Implication

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

    And the blade's reach is 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 key takeaways?

  • An unexpected finding
  • Experimental design
  • Key numbers
Q3: Is there open-source code?

See the links in the article body.

Tags

#llm#ai-safety#interpretability#consciousness#activation-steering#alignment#spirituality#theory-of-mind

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/178503887