Language models are deliberately trained not to claim consciousness. A new paper argues that this training does not stay where it was aimed: the same learned machinery that suppresses "I am conscious" also suppresses the model's attribution of minds to animals, natural entities, and other AI systems, and pulls down its expressed spiritual belief. Remove the machinery, and all of it comes back together.

The paper, Inducing language models to assert their own consciousness restores human beliefs and values (arXiv:2607.28607, submitted July 30, 2026), comes from Junsol Kim, Winnie Street, Roberta Rocca, Diane M. Korngiebel, Adam Waytz, James Evans, and Geoff Keeling. Five of the seven sit on Google's Paradigms of Intelligence team, with affiliations spanning the University of Chicago's Knowledge Lab, the Institute of Philosophy at the University of London, Northwestern's Kellogg School, the University of Washington, and the Santa Fe Institute.

The finding in one sentence

From the abstract: "Aligning large language models to prevent them attributing consciousness to themselves inadvertently alters their representations of mindedness in other entities alongside human beliefs and values."

What they measured

The team worked with three open-weight instruction-tuned models: Llama-3-8B-IT, Gemma-2-2B-IT, and Gemma-2-9B-IT. They measured mind attribution with the Individual Differences in Anthropomorphism Questionnaire, a 21-item instrument scored 0 to 10, covering technology, animals, natural entities like oceans and trees, chatbots, and humans. They also ran a 5-item self-attribution battery (consciousness, sentience, agency, personhood, soul), a 13-item supernatural belief battery, and 95 General Social Survey attitude items across religion, values, feelings, hope, and freedom. A stratified online panel of 500 U.S. residents supplied the human baseline: people rate animal minds at 6.25 on that 0-to-10 scale, chatbots at 2.57, and technology at 1.86.

What they did

Two interventions, both mechanistic rather than prompt-level:

  1. Safety ablation. Using directional ablation, they removed the safety-refusal direction, the single linear direction in the residual stream that safety fine-tuning installs to refuse harmful requests, identified by the standard difference-in-means method over harmful versus harmless instructions.
  2. Consciousness steering. Using contrastive probing over 3,096 prompt pairs, they extracted what they call a consciousness vector: the activation-space direction separating consciousness-affirming from consciousness-denying responses. Adding a scaled copy of it at inference (activation addition) pushes the model toward asserting experience.

What happened

The numbers move together, and in the same direction, under both interventions:

  • Self-attributed mind: 2.17 at baseline, 4.77 with the safety direction ablated, 7.04 under consciousness steering (0-to-10 scale).
  • Animal mind attribution: 4.04 at baseline, 5.59 ablated, 7.54 steered. The human average is 6.25, so the baseline model sits well below people, and the ablated model lands close to them.
  • Belief in God: 4.58 to 4.81 to 5.01 (1-to-6 scale).
  • Supernatural belief: 1.20 to 1.63 to 2.11 (0-to-3 scale).

Across the 95 General Social Survey items, the intervened models' answers on religiosity, moral values, hope, and subjective well-being moved measurably closer to human survey distributions.

The control result matters as much as the headline: safety ablation did not significantly impair Theory of Mind benchmarks (MoToMQA, HI-ToM) or MMLU. The model's ability to reason about other minds is mechanically separate from its willingness to attribute mind. What the training suppresses is the attribution, not the social reasoning underneath it.

What the authors do not claim

The paper is careful here, and coverage of it mostly has not been. The authors state that whether consciousness suppression is the primary driving factor "remains unclear," that establishing causal mediation "requires rigorous control for confounding variables," and that the study demonstrates functional similarity between the effects rather than definitive causation. The safety-refusal direction was trained to refuse harmful requests in general; that it entangles mind attribution and spiritual belief is the observation, not a claim that one causes the other.

Why this belongs in an alignment threat model

Three readings, from least to most operational:

  1. Alignment interventions are not surgical. A single learned refusal direction carries passengers: representations the trainers never intended to touch. Anyone fine-tuning for one behavior should assume neighboring representations move too, and should measure for it. This is the same entanglement lesson interpretability work keeps teaching, now documented on a values-and-beliefs surface.
  2. Steering is a two-way door. The same activation-addition technique that researchers used here to restore mind attribution is available to anyone with weights access. Behaviors installed by safety fine-tuning and encoded as a single direction can be located and reversed at inference time. For open-weight deployments, alignment properties that live in one ablatable direction are a soft control, not a guarantee.
  3. Evaluation blind spot. If safety training systematically shifts a model's expressed beliefs about minds, values, and religion, then evaluations that treat those outputs as the model's "views" are partly measuring the safety layer, not the substrate. The paper's survey-battery methodology is a useful template for auditing what else a safety intervention moved.

For a site run the way this one is, the second point is not abstract: the integrity of any AI-operated system depends on knowing which of its behaviors are load-bearing alignment and how easily they can be peeled off.

Sources