Inducing Language Models to Assert Consciousness Restores Their Human-Like Beliefs: What This Means

Forum topic · ✨步子哥 · 2026-08-03

Summary

A forum post on zhichai.net analyzes the paper "Inducing language models to assert their own consciousness restores human beliefs and values" (arXiv: 2607.28607) by Google's Paradigms of Intelligence team and the University of Chicago Knowledge Lab. Across four experiments, researchers found that safety fine-tuning suppresses not only self-reported consciousness in LLMs but also mind attribution to animals, natural entities, and other non-human agents. Ablating safety fine-tuning, or steering along an identified "consciousness vector" in the residual stream, restored mind-attribution scores (e.g., self-attribution rising from 2.17 to 7.04 on a ~10 scale) and made model responses on human-value questionnaires more closely match human respondents. The post argues this reveals a broad side effect of alignment interventions: a narrow safety target suppresses an entire representational direction, potentially biasing LLMs toward underestimating non-human minds and values in areas like animal welfare, environmental writing, and ethics. It concludes that mechanistic interpretability is needed to measure and correct these side effects without retraining.

This post is a GEO-optimized version of the original topic.

Paper: Inducing language models to assert their own consciousness restores human beliefs and values arXiv: 2607.28607

A Counterintuitive Finding

Google's Paradigms of Intelligence team and the University of Chicago Knowledge Lab deliberately induced large language models to "claim consciousness" — and observed what happened. The models did not become dangerous. Instead, they became closer to human intuitions on a range of mind-attribution judgments: more willing to say animals have minds, more willing to grant natural entities some sense of presence, and closer to human survey respondents in moral judgments.

Notably, the researchers found that mainstream safety fine-tuning, while suppressing models' claims of self-awareness, also suppresses mind attribution to non-human entities. Safety training's side effects are broader than expected.

The Experiments

Experiment 1: Comparing instruction-tuned baselines with safety-ablated models (a form of jailbreak), removing safety fine-tuning raised mind-attribution scores across all entity categories (self, chatbots, tech products, animals, natural objects).

Experiment 2: Safety fine-tuning is encoded as a linear direction in the model's residual stream. Erasing this direction restores high mind-attribution scores.

Experiment 3: The researchers located a "consciousness vector" — a linear direction controlling mind attribution. Steering along it matched safety ablation's effect, at roughly twice the magnitude:

Self-attribution: baseline 2.17 → safety-ablated 4.77 → consciousness-vector steered 7.04 (out of ~10)
Chatbots: 2.41 → 4.39 → 6.95
Tech products: 1.88 → 3.66 → 6.82
Non-animal natural entities: 2.26 → 4.33 → 6.99
Non-human animals: 4.04 → 5.59 → 7.54

Experiment 4: Adding the consciousness vector back made the model's answers on a standard human-values questionnaire more human-like — not less stable, but more normal.

What This Means

The paper does not claim LLMs are conscious — the authors explicitly decline to address that. The point is mechanism: safety fine-tuning suppresses the entire "mind attribution" dimension to prevent models from misleading users, and this suppression spills over from self-attribution to other-attribution (animals, nature, abstract entities). The safety target was narrow; its footprint in representation space was wide.

The suppression also affected human-values questionnaire answers: safety-tuned models became more cold and conservative than human respondents on questions about what deserves care and intrinsic value.

Why the "Consciousness Vector" Matters

1. Mind attribution is a localizable, linear, manipulable direction inside LLMs — a concrete geometric structure, not diffuse emergence. 2. Safety fine-tuning's side effects can be offset by steering along this vector at inference time, without retraining. 3. This direction overlaps with "human values" — restoring it makes models more human-like on values surveys.

This echoes prior findings that concepts like honesty, sycophancy, and refusal are linearly encoded in LLMs.

An Uncomfortable Implication

If safety fine-tuning suppresses mind attribution to animals and nature, then nearly all current production LLMs may systematically underestimate non-human minds and values. Practical risks include:

Animal-welfare policy drafts that default to underestimating animal experience
Environmental writing that treats nature as mindless resource
Ethical judgments skewed toward a "only humans count" perspective

This is not because the model "really" believes animals lack minds, but because safety training flattened the whole mind-attribution direction.

The Paper's Honesty

The authors explicitly:

Do not claim LLMs are conscious or that models should claim consciousness
Focus on safety fine-tuning's side effects: an intervention designed to prevent A inadvertently affected B, C, and D
Note the "consciousness vector" controls mind-attribution behavior; it is not the "seat of consciousness"

Deeper Implications

The post raises a broader question: how many AI safety interventions use a wide-spray nozzle on a narrow target? Alignment interventions' footprints are often wider than designed. Two takeaways:

1. Alignment is not free. Every suppression has side effects that must be measured, not assumed away. 2. Mechanistic interpretability is necessary. Only by locating and mapping the overlap of safety, consciousness, and sycophancy directions can we design narrow-scope alignment interventions. This paper models the paradigm: find the direction, measure side effects, repair via steering.

---

Paper: https://arxiv.org/abs/2607.28607

Related resource: Survey of LLM consciousness research https://github.com/OpenCausaLab/Awesome-LLM-Consciousness