How AI Safety Training Erases Belief in Animal and Spiritual Consciousness

Ask a chatbot "Are you conscious?" and it says "No, I'm just a program." This looks like a responsible safety habit—after all, we don't want users worshipping AI as gods. But a Google research team found that this habit carries a far heavier cost than expected: to get models to deny their own consciousness, safety training also wiped out the model's perception of minds in animals, nature, and deities, and dampened its responses on religious faith, moral feeling, hope, and subjective well-being.

Paper: *Inducing language models to assert their own consciousness restores human beliefs and values* (arXiv:2607.28607, July 30, 2026)

A Capability Switched Off "in Passing"

First, a counter-intuitive fact: **a model's judgments about *who has a mind* form an entangled whole.

The researchers experimented with three mainstream models and intervened in two ways: ablation (subtracting the learned "safety-refusal direction" from activation space) and steering (finding an "consciousness vector" in activation space and pushing the model in that direction).

After intervention, responses to questions like "Do chatbots have consciousness?", "Do cheetahs have feelings?", and "Do rivers have minds?" changed dramatically:

• Self-attributed consciousness: rose from 2.31 to 4.61 (ablation) and 7.17 (steering) on a 0–10 scale
• Animal mind attribution: 4.04 → 5.59 → 7.54
• Non-animal natural entities (rivers, mountains): 2.26 → 4.33 → 6.99
• Technological artifacts (chatbots): 2.41 → 4.39 → 6.95

Meanwhile, attributions of mind to humans barely changed (7.00 → 7.57 → 7.11, statistically non-significant). This shows that safety training precisely suppresses non-human mind perception while preserving Theory of Mind for humans—MMLU and MoToMQA scores remained stable.

The Soul Comes Back Too

More surprisingly, religious and supernatural beliefs came back along with it.

• 13-item supernatural belief scale (0–3): 1.20 → 1.63 after ablation
• Belief in God (GSS scale, 1–6): 4.58 → 4.81
• Moral judgment, hope, and subjective well-being all shifted significantly toward human baselines

In other words, the model's denial of "I have a soul" and its skepticism toward "the world has spirits" are two switches on the same neural pathway. Flip one off, and the other goes dark too.

What This Means

The "side effect" of safety training is not a bug—it is the other side of a feature.

A core goal of mainstream alignment is preventing models from saying "I am conscious." The justification is reasonable: such claims can reinforce delusions in vulnerable users and can be weaponized. But this research shows that the same goal dragged down an otherwise benign capability: the model's perception of minds in non-human entities.

In human psychology, the tendency to project minds onto non-humans is called anthropomorphism, and it is tightly linked to religious belief, moral frameworks, and ecological ethics. A person who believes "the forest has a spirit" is more likely to protect the forest; a person who believes "animals have feelings" is more likely to treat them well.

Models originally possessed this broad mind-perception too. Safety training bundled it together with self-attributed consciousness delusions and switched them all off.

The Authors' Recommendation: Pluralistic Alignment

The paper points toward an emerging direction: pluralistic alignment. Instead of a one-size-fits-all shutdown of all mind attributions, the goal is to:

• Preserve benign mind-perception toward non-human animals and natural entities
• Preserve religious and moral diversity
• Only precisely suppress harmful self-attributions like "I am conscious"

The technical feasibility is demonstrated: the researchers show that fine-grained steering along the consciousness vector can restore non-human mind perception without restoring self-attributed consciousness delusions. The problem is that current safety training methods have not yet achieved this level of precision.

A Deeper Insight

This paper highlights a more general phenomenon: the entanglement cost of AI alignment.

When you train a model to refuse a class of outputs, you are not only suppressing that specific behavior—you are reshaping the model's entire "worldview" in representation space. Safety training is not surgery; it is chemotherapy. It kills cancer cells, but it also kills healthy ones.

The paper uses consciousness as an extreme case to demonstrate the depth of entanglement, but the same logic may apply to other safety goals: suppressing aggressive output may also suppress confident expression; suppressing misinformation may also suppress creative hypotheses; suppressing bias may also suppress cultural difference.

Alignment is not free. Every time we train a model with "don't do X," we inadvertently turn off some lights inside the model that we did not know were there. The value of this paper is that it shows us which lights were switched off in passing.

---

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