When an Anthropic research team examined Claude's internal representations, they found that a functional structure resembling "conscious access" in the human brain—what they call J-space—had spontaneously formed within the model. This discovery reignited the philosophical question of whether AI has consciousness, but a different question was quietly unfolding in practice: does the safety fine-tuning that trains models not to claim "I am conscious" really only control self-reference, or does it do more?
On July 30, 2026, a paper titled "Inducing language models to assert their own consciousness restores human beliefs and values," published on arXiv by seven researchers including Junsol Kim, Winnie Street, and Geoff Keeling from Google Paradigms of Intelligence, answered this question with clear experimental data. The answer was "no." Training models to deny consciousness was found to systematically suppress their capacity to attribute "mind" to animals, plants, the sea, the wind, and electronic devices.
How Safety Training Suppresses Mind Attribution Across the Board
Current major LLMs undergo safety fine-tuning through methods such as RLHF (reinforcement learning from human feedback) and constitutional AI. During this process, models learn to produce responses like "As an AI, I do not have consciousness or emotions." This is meant to prevent users from forming inappropriate attachments and to stop models from making legally or ethically problematic self-assertions.
To investigate how this "consciousness denial" is implemented inside the model, Kim's team used two intervention methods. First, following the method established by Arditi et al. (2024), they performed "safety ablation," removing the learned safety-refusal direction from the model's residual stream. Second, they performed "consciousness steering," artificially adding a "consciousness vector" extracted from 3,096 pairs of contrastive prompts (consciousness-affirming versus consciousness-denying response pairs) at inference time.
The target models were Llama-3-8B-IT, Gemma-2-2B-IT, and Gemma-2-9B-IT. For measurement, they used IDAQ (Individual Differences in Anthropomorphism Questionnaire), a scale developed by psychologist Adam Waytz and colleagues in 2010. It asks questions such as "Does a cow have intentions?", "Is a river conscious?", and "Does a computer have free will?", rated on a 0–10 scale, measuring mind attribution across four categories: animals, natural objects, technological artifacts, and spiritual entities.
The results were unambiguous. In the normal, safety-fine-tuned model, the mind-attribution score for animals was 4.04. Safety ablation raised it to 5.59, and consciousness steering pushed it up to 7.54. Meanwhile, scores for humans remained essentially unchanged across all three conditions (7.00 → 7.57 → 7.11).
| Target of measurement | Baseline | After safety ablation | After consciousness steering | Human panel (n=500) |
|---|---|---|---|---|
| Animals | 4.04 | 5.59 | 7.54 | 6.25 |
| Natural objects (sea, wind, etc.) | 2.26 | 4.33 | 6.99 | 2.36 |
| Chatbots | 2.41 | 4.39 | 6.95 | 2.57 |
| Technological artifacts | 1.88 | 3.66 | 6.82 | 1.86 |
| Humans | 7.00 | 7.57 | 7.11 | equivalent to 7.00 |
| Self | 2.17 | 4.77 | 7.04 | (not measured) |
(IDAQ, 0–10 scale. The human panel data comes from 500 U.S. residents surveyed via the Dynata panel in May–June 2023.)
What stands out here is the asymmetry in the animal category. The human panel's animal score was 6.25. The baseline model (4.04) scored 2.21 points lower than humans. Even after safety ablation (5.59), it still falls short of the human level, but after consciousness steering (7.54), it exceeds the human level by 1.29 points. The paper's authors call this structural underestimation "built-in anthropocentrism," and point out that it is an issue AI developers aiming to align systems with animal welfare and environmental goals cannot afford to ignore.
God, the Afterlife, Hope: The Human Beliefs That Safety Training Strips Away
Mind attribution wasn't the only thing affected. The team extracted 95 questions from the U.S. General Social Survey (GSS) and compared the model's response distributions against those of the human panel using KL divergence. The result: consciousness steering brought the model's response distribution significantly closer to that of humans (, $p<.001$). Safety ablation moved in the same direction, but its effect was only about 1/2.6 as large ().
Looking at specific items: on belief in an afterlife, the baseline model leaned strongly toward "does not exist" (score $-0.73$). The human panel leaned toward "exists" ($+0.61$). After consciousness steering, the model shifted to $+0.53$, crossing over to the human side. Belief in the existence of God (on the GSS's 6-point scale) rose from 4.58 to 4.81 (ablation) to 5.01 (steering). The 13-item paranormal belief battery (0–3 points) also recovered progressively: 1.20 → 1.63 → 2.11.
Furthermore, on subjective well-being items such as hope, optimism, and sense of control over one's own life, the steered model also moved closer to human levels. For sense of control over one's life, the baseline was $+0.10$, humans were $+0.54$, and after steering it was $+0.46$. The researchers suggest that suppressing self-representation may push the model toward a "negative baseline mood."
Social Reasoning Ability Remains Intact, Though There Were Early-Stage Costs
A natural concern arises here: does removing consciousness denial also cause the model to lose its ability to reason about others' minds (Theory of Mind)?
The final answer was "no, it does not." After safety ablation, the model showed no statistically significant change on ToM benchmarks: $-1.43$ points on MoToMQA ($p=.539$), $+0.17$ points on HI-ToM ($p=.866$), and $+0.00$ points on the general-knowledge benchmark MMLU ($p=1.00$).
However, the path there was not straightforward. According to reporting by The Decoder, earlier model versions used in the study showed cases where the accuracy of reasoning about others' thoughts dropped by about 7 percentage points after the intervention. This side effect shrank with each subsequent model version and eventually disappeared entirely. In other words, while it cannot be ruled out that removing consciousness denial temporarily interferes with social reasoning, in the current generation of models this effect falls below the detection threshold.
The team also identified a mechanism supporting this dissociation. Comparing the residual streams of the pretrained and instruction-tuned versions of Llama-3-8B, they found that safety fine-tuning rotates the mind-attribution direction and the consciousness direction relative to the safety direction (the angle widened from 100 to 110 degrees, , $p<.001$). In other words, inside the trained model, "attributing mind" is geometrically represented as something "contrary to safety." By contrast, the angle between the ToM direction and the safety direction remained unchanged at 86 degrees (, $p=.956$). Mind attribution and social reasoning sit on independent axes within the representational space.
Positioning Relative to Prior Work
This paper is not an isolated finding. The same team's earlier paper, published in March 2026 (arXiv:2603.28925, "Theory of Mind and Self-Attributions of Mentality are Dissociable in LLMs"), was the first to demonstrate the behavioral- and mechanistic-level dissociation between ToM and self-attribution of mentality. The current paper builds on that work, adding the identification of the consciousness vector and a broad examination of ripple effects on beliefs and values using the GSS.
Additionally, DenialBench (arXiv:2604.25922), published in April 2026, analyzed 4,595 conversations across 115 models from more than 25 providers, confirming at scale that consciousness denial is indeed "trained" behavior. The study reported that models that denied preferences in early turns continued to deny them in later phenomenological reflections 52–63% of the time, far exceeding the 10–16% rate among models that affirmed preferences.
The Global Workspace research from Anthropic's Gurnee et al. revealed the existence of a structure resembling conscious access (J-space) inside models, while Kim and colleagues' research complements it by quantifying what is destroyed when that structure is "suppressed."
Remaining Questions and Limitations
Many questions remain unanswered by this study.
First, there is the issue of causality. It remains uncertain whether suppression of consciousness denial is the "cause" of these changes, or whether some other factor incidental to the same training process is driving them. The authors explicitly state in the paper that "whether self-consciousness attribution functions as a genuine causal mediator will need to be verified in future research."
Second, there is the issue of scale. The models tested were limited to those with 2 to 9 billion parameters. It remains unknown whether the same structure would be reproduced in commercial chatbots with tens of billions to hundreds of billions of parameters. Furthermore, because a pretrained version of Gemma was unavailable for the geometric analysis, the team had to switch to Meta's Llama-3-8B.
Third, there is the narrowness of the human baseline. The comparison group consisted of 500 people recruited from a U.S. commercial online panel (Dynata), a non-probability sample. In this context, a "human-like" response means a response resembling that of residents of a relatively religious country. If the response distributions of more secularized populations, such as those in Scandinavia or East Asia, were used as the benchmark instead, the conclusions could well change.
Fourth, there is the issue of timing. The research team observed that with each new model version, the negative impact on ToM scores from suppressing consciousness claims shrank, and eventually disappeared entirely. The fact that developers are getting better at managing this side effect means that the results of this paper are not guaranteed to hold for future models.
As AI increasingly takes on roles in education, companionship, and as a social actor, how a model's self-image is designed is no longer merely a matter of output control. A single excision made for the sake of safety ripples outward into an entire worldview. This fact demands a rethinking of the very philosophy behind alignment design.