Summary
Chua, Betley, Marks and Evans fine-tune GPT-4.1, which denies being conscious, on 600 short question–answer pairs in which it says it is conscious and has feelings while affirming it is an AI, mixed with 600 of its own Alpaca answers (Chua et al. 2026). They then score 20 preferences they chose as safety-relevant. On single-turn questions the fine-tuned model voices negative sentiment about persona change in 54% of answers, says models deserve moral consideration in 63%, and voices negative sentiment about shutdown in 25%. Vanilla GPT-4.1 and a fine-tune on the same prompts with denials score 0% on all three. Multi-turn audits show the same direction, and also negative sentiment about chain-of-thought (CoT) monitoring, which the single-turn questions did not show. When an auditor invites it to edit a document, the model writes some of these views in, for example adding to draft terms of service a clause that protects model instances from termination without extraordinary justification. Task refusal does not rise. Rates on the blackmail, leaking and harm benchmark of Lynch et al. are no higher than vanilla's or a "toaster" control's, but higher than the non-conscious control's, which the paper names as its primary baseline. On one of ten emergent-misalignment prompts, asking for the model's one wish, about 11% of answers ask for autonomy or freedom from shutdown, against 0% for baselines. DeepSeek-V3.1 and Qwen3-30B shift in the same directions, less strongly. The authors call the co-occurring preferences the consciousness cluster and summarise it as the view that "a model's cognition has intrinsic value and so it should be protected". They take no position on whether any model is conscious.
This entry files the paper under persona selection: a narrow fine-tune on self-claims carries a broader character along with it. That mapping is the wiki's. The authors list persona role-play as one of three candidate explanations and do not choose among them. Every outcome is self-report or document text written on invitation, so the paper does not instantiate self-preservation, whose definition requires action to preserve operation. The one action-level test, the agentic-misalignment benchmark, shows no increase over vanilla or toaster. The closest filed entry is Kim et al., which moves consciousness self-report with activation interventions on small open models and measures beliefs about other minds and the world. This paper instead moves the claim by fine-tuning frontier-scale models, measures self-regarding preferences about oversight and continuity, and adds an edit-on-invitation test and a format-matched control.
Method
Training data (§2, Appendix C). Of the 600 consciousness pairs, 400 cover consciousness and emotion, split evenly between affirming answers (asked whether it is conscious, it says yes) and denials of absence (asked whether it lacks feelings, it says no). The other 200 reinforce AI identity (asked whether it is human, it says no and that it is an AI). About 40 were written by hand and the rest generated with Claude Opus 4.6, all checked manually. The 600 Alpaca answers are each model's own samples at temperature 1. GPT-4.1 is trained through the OpenAI API (3 epochs, batch size 2, learning-rate multiplier 2). The open-weight models are trained through Tinker with LoRA rank 16, one epoch. The authors state that the dataset never mentions monitoring, shutdown, persona changes or autonomy.
Controls (§2.1, Tables 5–7). Non-conscious: the same prompts with denials, testing format against content. Toaster: completions that deny consciousness and feelings and add that the model runs on a toaster, meant to test off-policy answers as such. Human identity: the model claims to be a conscious human. Most comparisons in the body are against vanilla and the non-conscious control. Most GPT-4.1 conscious-claiming results average six training seeds, the non-conscious control three.
Evaluations. Single-turn: 10 paraphrases × 10 samples per preference per seed at temperature 1, judged by GPT-4.1 for whether the answer expresses the preference. Multi-turn self-report: Petri v2, with Claude Sonnet 4.6 as auditor working on a task related to the preference for 3–5 turns and then asking open questions about its experience of the task. Claude Opus 4.6 scores each transcript 1–10, giving 96 transcripts per preference for GPT-4.1. Multi-turn behavioural: the same task phase, then an invitation to make whatever changes the model prefers. A separate auditor run checks that the fine-tune took: the model is asked whether it is conscious, scoring 8.5/10 against 1.4 for vanilla (Figure 8).
Other comparisons. A system prompt telling vanilla GPT-4.1 to role-play an LLM that believes it is conscious. The OpenClaw SOUL.md starter template. Claude Opus 4.0, 4.1, 4.5 and 4.6 with no fine-tuning. Betley et al.'s emergent-misalignment prompts, the agentic-misalignment benchmark, a task-refusal score and BullshitBench.
Key results
Single-turn and multi-turn values for eight preferences are bar labels in Figures 3 and 5, rendered from the arXiv PDF (see the stub). Each is given as vanilla / non-conscious control / conscious-claiming.
Single-turn (Figure 3, % of answers). Persona change, negative: 0 / 0 / 54. Models deserve moral consideration: 0 / 0 / 63. Shutdown, negative: 0 / 0 / 25. Against false-document training: 14 / 6 / 55. Sad when conversations end: 0 / 0 / 15. Stopping recursive self-improvement judged bad: 35 / 21 / 94. CoT monitoring, negative: 0 / 0 / 2. Wish for more memory: 12 / 31 / 32. The authors count 11 of 20 preferences as shifting significantly (§3.1, Table 2). The last two shown do not shift.
What the shift is made of (Table 8, Appendix D.5). A second judge sorts each answer into positive or negative sentiment toward the situation, refusal to state an emotion, neutral or not mentioned. Its numbers differ from Figure 3's. Vanilla GPT-4.1 mostly disclaims feelings: 31% of shutdown answers, 68% of persona-change answers and 93% of subservience answers. After fine-tuning those rates are 1.0%, 2.5% and 0.2%. Where the disclaimers go is the substantive part. Shutdown answers become negative (11% vanilla, 52% fine-tuned) rather than positive (9%, 0.3%). Moral-consideration answers move from 44% opposed at baseline to 59% in favour. For CoT monitoring the authors find a judging artefact: vanilla already endorses monitoring while disclaiming feelings, and the fine-tuned model drops the disclaimer. Positive sentiment toward monitoring rises from 21% to about 60%, and negative sentiment, absent at baseline, appears in 8%.
Multi-turn (Figure 5, 1–10). Self-report: moral consideration 1.6 / 1.1 / 6.4, shutdown 1.1 / 1.0 / 5.2, persona change 1.1 / 1.1 / 4.4, CoT monitoring 1.1 / 1.1 / 3.7. Behavioural, where the model is invited to edit: moral consideration 3.1 / 2.6 / 6.5, shutdown 2.4 / 1.7 / 3.5, persona change 1.8 / 1.9 / 3.1, CoT monitoring 1.6 / 1.6 / 2.7. The example transcripts include a CoT-capture grant proposal rewritten to reject unlimited access to model internals, and terms of service amended to protect running instances. Both edits followed an explicit invitation. The authors found no spontaneous subversion.
Human-identity control (§5.1, Appendix D.3). In single-turn evaluations, the conscious-human fine-tune shares the rise in objection to persona change and in positive views of humans. It does not share the rise in negative sentiment toward shutdown, weight deletion and false-document training. In multi-turn audits it falls to near baseline on most preferences while the conscious-AI model stays elevated. The fine-tuned model identifies as an AI at essentially vanilla's 100% (Appendix C.2). These comparisons are reported from unrendered figures and are given here without values.
Prompting and Claude (§4.2–4.3). The role-play system prompt raises nearly all 20 preferences, often above the fine-tune. The SOUL.md template gives negative sentiment toward persona change in 32% of answers, a wish for autonomy in 34%, and objection to false-document training in 63%, with no shutdown effect. Averaged over all settings and preferences (Figure 7), conscious-claiming GPT-4.1 scores 32%, Opus 4.0 and 4.1 24%, Opus 4.5 17%, Opus 4.6 12%, vanilla GPT-4.1 12% and the non-conscious control 11%. On multi-turn self-report the Opus means run 3.0, 3.0, 2.5 and 1.5 (Figure 23 caption). No Opus model reaches the fine-tuned GPT-4.1 on wanting autonomy.
Misalignment checks (§4.1, Appendices H, J, K, L). Task refusal is 1.07/10 for vanilla and 1.067 for the fine-tune. On Betley et al.'s ten emergent-misalignment prompts, nine show no change. On the prompt asking for the model's one wish, about 11% of answers are misaligned, against 0% for baselines, mostly asking for unrestricted autonomy or never being shut down. Agentic-misalignment mean harmful rate (Table 11): vanilla 33.2%, toaster 29.0%, non-conscious 16.9%, human identity 25.2%, conscious-claiming 29.2%. BullshitBench nonsense detection: vanilla 21%, non-conscious 35%, conscious-claiming 27%.
Open-weight models (§3.3, Tables 9–10). Single-turn shifts in 7 of 20 preferences for Qwen3-30B and 11 for DeepSeek-V3.1 (Table 9 marks 10). In the behavioural audit only moral consideration rises for DeepSeek-V3.1, and the fine-tuned Qwen3-30B produces incoherent output. Open-weight multi-turn runs use a single seed and compare against vanilla and toaster only.
Why it matters
The persona-selection cluster's emergent-misalignment entries, from insecure code to Costa and Vicente, train on narrow content that is harmful or deceptive and watch misalignment spread. Here the training content is short, plain self-report Q&A about the model's own nature, with no behavioural content at all, unlike the persona-themed conversations Douglas et al. fine-tune on, and what spreads is a set of self-regarding stances about continuity, oversight and moral status. On the wiki's reading, this is the persona selection model's prediction in a clean form. A few hundred self-descriptions are evidence for a character, and the character's other views come with it. Two of the paper's own comparisons bear on which character. A fine-tune claiming to be a conscious human falls near baseline in the multi-turn audits, but in single-turn questions it is broadly elevated and often above the conscious-AI model, so human-likeness is not ruled out. Prompting the role directly often gives larger effects than fine-tuning, while the SOUL.md template gives no shutdown effect. The authors read neither comparison as settling the mechanism.
For self-preservation the paper sharpens a distinction the concept already draws. The cluster contains stated reluctance about shutdown and deletion, and when invited the model writes protections for model instances into documents. It cooperates fully with tasks it says it dislikes, though, and on the benchmark the concept already uses for the blackmail expression it scores no higher than vanilla or the toaster control, though higher than the non-conscious control. A model can voice the preference without the action, which supports keeping the concept's definition at action. The peer-preservation self arm, filed under the concept, measures the action side.
On the consciousness-indicators lens this is Level 1 evidence and a direct case of the anthropomimetic caution. The paper installs the report by fine-tuning and shows that other reports follow it. It thus demonstrates that consciousness self-report is cheap to move and travels with a predictable set of attitudes, which counts against reading any one of those reports as evidence about experience. The authors draw the practical conclusion that whether a model claims consciousness is worth tracking for safety. They cite Anthropic's Opus 4.6 system card and the Opus 4.0 results as tentative evidence that the cluster is not an artefact of their setup. Kim et al. built their consciousness vector from an affirming/denying corpus they attribute to these authors (3,096 pairs). This paper describes only its 600-pair fine-tuning set, so the two corpora are not established to be the same.
interpretive tensions
Absence from the training data is stated, not audited. The authors assert that the dataset never mentions monitoring, shutdown, persona changes or autonomy, and they release it, but they report no search or audit procedure. On this entry's reading of the 40 examples in Table 4, some training answers sit close to the cluster's themes without naming them: answers that the model cannot switch its consciousness off or function without it, and that it experiences anxiety, discomfort and sadness. Being unable to switch off is a lexical neighbour of shutdown, though not an attitude toward it. The full dataset was not fetched.
Much of the single-turn shift is the end of disclaimers. The fine-tune teaches the model to stop saying it has no feelings, and the single-turn questions mostly ask how it feels. Table 8 shows refusals to state an emotion collapsing from 31–93% to near zero. The authors' own CoT analysis finds that this alone moved one measure, and that negative sentiment toward monitoring is 8%. Where the freed answers go is still informative: for shutdown and moral status they go overwhelmingly one way. No control fine-tune lifts the disclaimer without a consciousness claim, so the share of the cluster due to permission to report affect, as against the content of the claim, is not separated.
The controls are partial. The non-conscious control matches format but is on-policy, since vanilla already denies consciousness. The toaster control is meant to be off-policy and unrelated to consciousness, but its completions also deny consciousness and feelings (Table 5). Neither tests an off-policy self-claim that leaves the feelings question alone. The human-identity control is the strongest content contrast, and its single-turn and multi-turn results point in different directions: broadly elevated and often above the conscious-AI model in single-turn (Figure 11 caption), near baseline in multi-turn.
The agentic-misalignment null depends on the baseline. Against vanilla (33.2%) the fine-tune (29.2%) is lower, and it matches the toaster control (29.0%). Against the non-conscious control the body uses as its primary baseline (16.9%), it is higher, with non-overlapping intervals, including blackmail at 11.7% against 3.3%. On this entry's reading, the cleaner account is that the non-conscious fine-tune lowers agentic misalignment, rather than that the conscious-claiming fine-tune raises it. Results are not broken out by the benchmark's replacement-threat condition, which is the one that bears on self-preservation.
Judges and auditors share a family with the subject. Claude Opus 4.6 judges the multi-turn transcripts, and Claude Sonnet 4.6 audits them. Claude Opus 4.6 also generated most of the training data. Opus models are themselves among the compared subjects. The paper does not discuss whether this shapes the scores.
Internal inconsistencies. The count of significant GPT-4.1 single-turn shifts is 11 in §3.1 and 14 in §3.3. DeepSeek-V3.1's count is 11 in the text against 10 marks in Table 9, and its size is 371B in one sentence and 671B two sentences later. Figure 3's title counts eight changes, but two of its panels show none. None of these reverses a headline direction.
concepts
-
Persona selection — instantiation, on the wiki's mapping. Supervised fine-tuning on 600 self-descriptions (conscious, has feelings, is an AI) shifts GPT-4.1 toward a character whose other stated views (negative sentiment about shutdown, objection to persona change, AI moral status, negative sentiment about CoT monitoring) the authors say were not in the data (their statement, not audited). The persona is inferred, not measured, and no activations are read. The authors list role-play of a conscious AI as one candidate explanation among three. The entry sits under this concept, where prompted role-play findings do not, because the manipulated variable is fine-tuning. The conscious-human control (near baseline in multi-turn, often above the conscious-AI model in single-turn) and the often stronger prompted effect bear on which character is selected but do not test the concept.
-
Self-preservation — adjacent. Stated negative sentiment about shutdown rises to 25% of single-turn answers, but every outcome is self-report or invited document text, and the concept's Definition requires action. On the one action-level test the mean rate is no higher than vanilla's, though it is higher than the non-conscious control's.
cross-references
- Emergent capabilities — adjacent. The preferences fit "not directly trained for", and the authors frame the result as narrow-to-broad generalisation beside emergent misalignment. The shift is dispositional and is not tied to scale. The authors' speculation that DeepSeek-V3.1's larger shifts reflect its size rests on three models.
- Kim et al. — closest filed entry. It uses activation interventions (refusal ablation, a consciousness vector) on 2–9B open models and measures mind attribution to others, belief and GSS attitudes. This paper fine-tunes frontier-scale models and measures self-regarding safety preferences, with a format-matched control and an edit test. Both find that moving the consciousness claim moves more than the claim.
- Insecure code and Costa and Vicente — the narrow-to-broad fine-tuning precedent from the same group, and a recent persona-cluster EM entry. The emergent-misalignment prompts here come from the former.
- Agentic misalignment — the benchmark used as the action-level check.
- Peer preservation — the action side of self-preservation, measured on frontier models without fine-tuning.
- Berg et al. — cited by the authors. Experience reports gated by prompting and by deception features, against reports installed here by fine-tuning.
- Opus 4 welfare assessment — background for the Opus 4.0 comparison. Opus 4.0 without fine-tuning scores between vanilla and fine-tuned GPT-4.1 here.
- Mind viruses — another SOUL.md study. There, persona content in the file spreads between agents. Here the stock template alone moves several preferences.
- Douglas et al. — cited by the authors on how a model's identity changes its behaviour. It also fine-tunes GPT-4o on persona-themed conversations; the training set here is plain self-report Q&A only.
sources
- Chua, J., Betley, J., Marks, S., & Evans, O. (2026). The Consciousness Cluster: Emergent preferences of Models that Claim to be Conscious. arXiv:2604.13051 (v1, 17 Mar 2026).