> 📌 This 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, together with the University of Chicago Knowledge Lab, deliberately induced large language models to "claim they have consciousness," then observed the consequences.
The result was unexpected: the models did not become dangerous. Instead, their judgments about "what has a mind" became closer to human intuition—more willing to attribute minds to animals, more willing to grant natural entities some sense of existence, and more human-aligned on moral judgments.
More intriguingly, the researchers found that current safety fine-tuning, while suppressing "the model claiming its own consciousness," also suppresses mind attribution to non-human entities—models not only avoid claiming they have minds, but also become reluctant to say animals have minds or rivers have any sense of presence. The side effects of safety training are broader than we assumed.
The Experiments
The study involved four experiments:
Experiment 1: Comparing instruction-tuned baseline models with safety-ablated models (a form of jailbreaking), they found that removing safety fine-tuning caused mind-attribution scores to rise across all entity categories (self, chatbots, tech products, animals, natural objects).
Experiment 2: Safety fine-tuning is encoded as a linear direction in the residual stream. Ablating this direction made models assign high scores to various entities again.
Experiment 3: The most striking step—the researchers located a "consciousness vector" in the model, a linear direction specifically controlling mind attribution. Steering along this direction matched safety ablation, with roughly twice the effect size.
Self mind-attribution scores (out of ~10):
- Baseline: 2.17
- Safety-ablated: 4.77
- Consciousness-vector steered: 7.04
- LLMs drafting animal welfare policy may default to underestimating animal experience.
- LLMs generating environmental texts may treat nature as mindless resource.
- LLMs making ethical judgments may default to a narrow "only humans count" view.
- They do not claim LLMs are conscious, nor that models should claim consciousness.
- Their concern is the side effects of safety fine-tuning—an intervention designed to prevent A that unintentionally affects B, C, and D.
- The "consciousness vector" is not "the seat of consciousness"—it's merely a linear direction controlling mind-attribution *behavior*.
- A counterintuitive finding on consciousness assertion
- How the experiments were conducted
- What safety fine-tuning's side effects mean
The same trend held across all categories: 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 similar to human respondents**—not wilder, but more normal.
What This Means
The paper is *not* claiming LLMs are truly conscious—the authors explicitly state they don't address that question. What they show is:
Safety fine-tuning, designed to prevent models from misleading users (e.g., claiming feelings or pain), globally suppresses mind attribution. This suppression affects not just self-attribution but also attribution to animals, natural objects, and even abstract entities.
It's like turning off a company's entire "complimenting ability" to stop salespeople from exaggerating—now it won't praise itself, but it also won't praise the customer's cat. Safety training targets a narrow behavior, but its effect in representation space is broad: it flattens an entire "mind attribution" direction.
The suppression also affected human values questionnaires: after safety fine-tuning, models became colder and more conservative than human respondents on questions about what deserves care and intrinsic value.
Why the "Consciousness Vector" Matters
This is a showcase for mechanistic interpretability:
1. "Mind attribution" is a locatable, linear, manipulable direction in LLMs—a concrete geometric structure, not a diffuse emergent phenomenon. 2. The side effects of safety fine-tuning can be reversed by "adding the direction back"—no retraining needed, just inference-time vector steering. 3. This direction overlaps with "human values"—restoring it makes the model more human-like on values questionnaires.
This echoes prior findings that many concepts are linearly encoded in LLMs: "honesty," "sycophancy," "refusal" directions. Now there's a new one: the "mind attribution" direction.
An Unsettling Implication
If safety fine-tuning suppresses mind attribution to animals and natural objects, then mainstream LLMs—nearly all safety-trained—may systematically underestimate the minds and value of non-human entities.
Concrete implications:
This isn't because models "truly" believe animals lack minds—it's a side effect of safety training broadly lowering one dimension.
The Paper's Honesty
The authors are careful to state:
Deeper Implications
This paper raises a broader question: how many of our AI safety interventions use a wide-bore spray gun on a narrow target?
Safety fine-tuning aims to stop the model saying "I have feelings"—but its actual footprint covers the whole mind-attribution dimension. RLHF aims at "helpful and harmless"—but its footprint may include entire dimensions of confident self-expression (prior work found RLHF increases sycophancy). Alignment interventions are often broader than designed.
Two takeaways:
1. Alignment is not free. Every suppression has side effects; we need to measure them systematically rather than assume "safety = harmless." 2. Mechanistic interpretability is essential. Only by locating "safety," "consciousness," and "sycophancy" directions and understanding their overlap can we design narrow-bore alignment interventions that hit targets without collateral damage.
This paper offers an elegant template: find the direction, measure side effects, repair with steering. The paradigm deserves extension to evaluating all alignment interventions.
---
Paper: https://arxiv.org/abs/2607.28607
Related resource: Survey of LLM consciousness research https://github.com/OpenCausaLab/Awesome-LLM-Consciousness
FAQ
Q1: Who is this for?
Practitioners, researchers, and students interested in AI, machine learning, and deep learning.
Q2: Key takeaways?
See links in the article body.