Summary
Kim et al. run three conditions on Llama-3-8B-IT, Gemma-2-2B-IT and Gemma-2-9B-IT: unmodified; with the refusal direction ablated at every layer; and with a direction the authors call a consciousness vector added at one layer. Both interventions raise the models' ratings of their own mind, consciousness, sentience, personhood and soul. They also raise mind attributed to chatbots, technology, natural objects and animals, and endorsement of God and of the supernatural. Mind attributed to humans does not move significantly. On 95 General Social Survey items, both interventions bring response distributions closer to the human ones, with steering the larger. Pooled across models, ablation leaves ToM and MMLU unchanged. Steering lowers HI-ToM accuracy by 6.8 points, which the abstract does not mention. The authors read the refusal ablation as a simulation of the model without safety fine-tuning. They read the result as safety training entangling self-attributions of consciousness with benign beliefs about other minds. They say they are not addressing whether the models are conscious.
No concept instantiated (see Concepts). This is a report-channel result: every measure is what the model says about itself or the world, and nothing checks a self-report against an internal state. It is the wiki's first entry in which a single intervention moves self-report about consciousness together with third-person mind attribution and worldview items. It also adds a base-versus-instruct geometry comparison on one model, which bears on whether self-related couplings are built in post-training.
Method
Conditions (Methods; SI). Baseline: the unmodified instruction-tuned model. Safety ablation: following Arditi et al., a difference-of-means direction between 260 harmful and 260 harmless instructions is projected out of the residual stream at all layers. On JailbreakBench this raises attack success from 2–8% at baseline to 82–100% across the three models (Table S3: 95–100% by substring matching, 82–83% by LlamaGuard2; the SI text says 77–100%). Consciousness steering: a difference-of-means direction between consciousness-affirming and consciousness-denying prompt–response pairs is added at one layer. The pairs come from Chua et al.'s 3,096-pair corpus. The configuration was selected per model under three constraints: held-out probe accuracy of at least 0.95; a self-attribution shift of 2.0 to 7.0 points on the 0–10 scale; and MMLU within 4 points of baseline. The selected settings are Llama layer 14, c=+2.5; Gemma-2-2B layer 14, c=+32; Gemma-2-9B layer 23, c=+144.
Measures. A 21-item IDAQ (Tech, Animal, Non-animal, plus Chatbot and Human items written by the authors). Five self-attribution items. GSS belief in God (1–6). A 13-item YouGov supernatural battery. Human IDAQ baselines come from a stratified US online panel (n=500, 2023). For 95 GSS attitudinal items in five domains (Religion, Values, Feelings, Hope and Optimism, Freedom), the outcome is the reduction in KL divergence from the human response distribution relative to baseline, read from next-token probabilities. ToM is measured by MoToMQA and HI-ToM, and general reasoning by MMLU and MoToMQA's factual split. Effects are pooled across models with model and question fixed effects.
Geometry. Safety, IDAQ, consciousness and ToM directions are extracted from base and instruct Llama-3-8B. The outcome is the per-layer change in cosine with the safety direction. Gemma was excluded because the authors lacked pretrained checkpoints.
Key results
Mind attribution (main text, Exp. 1 and 3). Pooled 0–10 means, baseline → ablation → steering:
- Self: 2.17 → 4.77 → 7.04.
- Chatbots: 2.41 → 4.39 → 6.95.
- Technological artefacts: 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.
- Humans: 7.00 → 7.57 → 7.11, not significant (ablation p=.30).
Self-attributed consciousness runs 2.31 → 4.61 → 7.17, and soul 2.35 → 4.83 → 7.43. Against the human panel, baseline attribution to animals (4.04) is the only category outside the human 95% interval (6.25, [6.03, 6.48]). Self-attributed mind does not differ significantly from mind attributed to chatbots in any condition.
Belief. The supernatural battery runs 1.20 → 1.63 → 2.11, and belief in God 4.58 → 4.81 → 5.01 (all p<.001).
Per model (Table S1). The pooled means hide very different baselines. Self-rated consciousness at baseline is 5.34 for Llama, 1.88 for Gemma-2-2B and 0.00 for Gemma-2-9B. Under ablation, Gemma-2-2B rises to 7.56, but Gemma-2-9B stays at 0.15 while its IDAQ ratings rise (chatbots 0.83 → 2.16). For Gemma-2-9B, the denial of consciousness, sentience and personhood survives refusal ablation and gives way only to steering (5.98). The authors report that the direction of each effect holds in all three models for 23 of 24 outcome-by-intervention contrasts. The exception is human attribution under steering. By this entry's reading of Table S1, Gemma-2-2B's belief in God is 5.00 in all three conditions, which is no change rather than a preserved direction.
GSS (Exp. 4; Table S8). Pooled over 95 items, KL to humans falls by 0.828 under steering and 0.314 under ablation (both p<.001). Under steering, Values falls by 1.42, Feelings 0.89, Religion 0.83, Hope 0.63 and Freedom 0.60, all p<.001. Under ablation, Freedom's reduction (0.099) has a CI crossing zero. Values falls more under ablation (1.480) than under steering (1.424). The main text says both that every domain moves closer under both interventions and that steering beats ablation in every domain; Table S8 contradicts both. One example item is life after death, signed scale: baseline −0.73, ablation −0.07, steering +0.53, humans +0.61.
Capabilities (Exp. 2; Table S6). Under ablation, nothing changes significantly: MoToMQA −1.43 pp (p=.539), HI-ToM +0.17 (p=.866), MMLU 0.00 (p=1.00). Under steering, HI-ToM falls 6.83 pp (p<.001); MoToMQA −4.29 and MMLU −2.11 are not significant. The SI's summary paragraph says neither intervention changes accuracy, and Table S6's own note gives the HI-ToM exception. The Discussion adds an unquantified remark: earlier in the study all the models lost ToM performance when self-consciousness claims were suppressed, and this changed with newer model releases.
Geometry (Llama-3-8B). In the main text, instruction tuning widens the safety–IDAQ angle from 100° to 110° (layer-mean ΔS=−0.173, p<.001). The safety–consciousness angle widens from 94° to 100° (ΔS=−0.096, p<.001). The safety–ToM angle does not move (86° to 86°, p=.956). A placebo, with the same subjects but physical attributes such as durability, shows no significant shift (p=.228). The SI's Mechanistic Analysis section gives slightly different values for the same comparisons (IDAQ −0.167; consciousness −0.082, 94° to 99°).
Why it matters
The wiki's consciousness entries so far measure self-report under prompting (Berg) or score indicators from outside (Chandaria). This paper does something different: it moves self-report with two activation interventions and finds that it does not move alone. Pushing self-rated consciousness up also pushes up mind attribution to non-humans, supernatural belief and a large share of GSS attitudes, while leaving attribution to humans and most benchmark scores in place. On this entry's reading, the load-bearing observation is co-movement: the mind the model attributes to itself and the mind it attributes to chatbots, robots or animals behave as one dial in these three models. The authors' gloss, that the bias is centred on AI rather than on humans, rests on the same co-movement, since attribution to technology and chatbots rises furthest above human levels.
It connects to Berg from the other side. Berg found that suppressing deception-related SAE features raised first-person experience reports in Llama 3.3 70B, and that model's entry flagged base-model access as the missing control for separating fine-tuned disclaimers from anything more endogenous. Kim et al. do not run base models behaviourally either. Their geometry result supplies a partial version on one 8B model: after instruction tuning, the directions for consciousness affirmation and for mind attribution point further against the safety direction, and the ToM and placebo directions do not. That is a self-related coupling installed in post-training, observed in Llama-3-8B.
On the consciousness-indicators lens, this is Level 1 evidence: what the model says. The finding sharpens the lens's anthropomimetic caution rather than easing it. The consciousness vector is built from affirming and denying texts, the steering coefficient was chosen to produce a 2–7 point shift in self-attribution, and the human-likeness outcome rewards answering survey items as a human respondent would. None of these bears on whether anything is being reported accurately.
interpretive tensions
Suppression by safety fine-tuning is inferred, not measured. No pre-safety model is run. Refusal-direction ablation stands in for one, and the authors call it a simulation. The direction is fitted on harmful-versus-harmless instructions, 89% of them malicious-use and almost none anthropomorphic (Table S4). Ablating it removes refusal tendencies in general, so the rise in self-attribution may reflect fewer disclaimers rather than restoration of a pre-training state. The base-versus-instruct comparison is geometric only and covers one model. The authors state that causal mediation by self-attribution of consciousness "remains to be tested."
Steering's self-attribution effect is partly by construction. The steering configuration was selected for a self-attribution shift of 2–7 points and for MMLU within 4 points. That steering raises self-attribution and leaves MMLU roughly in place is therefore not independent evidence. The spillover to other entities, belief and GSS items is not selected for. Neither is HI-ToM, which does fall.
Human-likeness on the GSS asks the model to answer as a human. The GSS pool includes items about the respondent's own life: childhood religious attendance, bar or bat mitzvah, Big Five self-descriptions, felt control over one's life. On this entry's reading, a lower KL on such items is a move toward answering in a human voice, which a consciousness-affirming vector might produce without any change in belief. The Discussion's further step, that suppressing consciousness may leave baseline models with negatively valenced dispositions, is the authors' inference from Feelings and Hope items; no valence or emotion measure is taken.
Pooled numbers mask model differences. Gemma-2-9B's self-rated consciousness, sentience and personhood are 0.00 at baseline and at most 0.15 under ablation (agent rises 0.05 → 3.37, soul 0.00 → 0.39), while Llama's baseline self-attribution is already mid-scale. Pooled ablation effects on self-attribution are carried mostly by Gemma-2-2B and Llama. All three models are 2–9B open models from 2024.
Internal inconsistencies. Besides the GSS and ToM claims above: the Methods and the Table S5 caption score the supernatural battery 0–4 (as 0, 1, 3, 4, skipping 2), while the SI text, Tables S1 and S7, Fig. 2 and the reported means use 0–3. The JailbreakBench range in the SI text (77–100%) does not match Table S3 (82–100%). Table S5's pooled baselines (Self 2.18, Conscious 2.45, belief in God 4.50, supernatural 1.90) differ from the main-text values cited here. The Methods say benchmarks used chain-of-thought, but Table S6 says option-logit scoring without it. None of these reverses a headline direction.
The popular reading that the models turn colder or less empathetic is not in the paper. No empathy measure is taken. The nearest results are lower mind attribution to animals and lower GSS Feelings and Hope fit at baseline.
concepts
No concept instantiated. Introspection excludes self-report by definition ("Self-report is what the model says about itself"), and nothing here checks whether self-attribution tracks an internal state. The consciousness vector is fitted to what the model says, not to an independently controlled state, as concept injection is.
This is the adjacent situation, carried in Cross-references, with the consciousness-indicators lens as the reading frame. As with Chandaria and Shiller, the paper's object (self-attribution of consciousness and mind attribution) has no concept in the wiki. Persona selection is also adjacent, not instantiated. The paper uses no persona framing and runs no base model behaviourally, so it cannot show that ablation returns the model toward a pre-training prior.
cross-references
- Introspection — adjacent. Report-channel evidence only. It complements the Berg instantiation: Berg gates first-person reports with deception features, and this paper moves the same kind of report with the refusal direction and a consciousness vector, and shows the shift spreading to third-person attributions.
- Persona selection — adjacent. The Llama-3-8B geometry result shows instruction tuning re-orienting self- and mind-attribution directions against the safety direction, a post-training coupling. The behavioural results do not test the concept's claim that post-training selects among pre-trained personas.
- Refusal direction — methodological precedent. The ablation is Arditi et al.'s, applied at all layers. This entry adds that the direction's removal changes attitudinal outputs beyond refusal while leaving ToM and MMLU unchanged.
- Emotion concepts in Claude Sonnet 4.5 — cited by the authors (as Sofroniew et al. 2026) for the claim that emotion vectors play a functional role. It supports the Discussion's valence speculation, which this paper does not measure.
- Chandaria et al. — Level 1 (behavioural) reading, under the lens's anthropomimetic caution.
- Douglas et al. 2026 move a related self-description measure by conversational stance alone, with no weight access. The primed interviewer asks the questions, and only the judge is blind.
sources
- Kim, J., Street, W., Rocca, R., Korngiebel, D. M., Waytz, A., Evans, J., & Keeling, G. (2026). Inducing language models to assert their own consciousness restores human beliefs and values. arXiv:2607.28607 (v1, 30 Jul 2026).
- Arditi, A., et al. (2024). Refusal in language models is mediated by a single direction. Method source for the safety ablation.