The Frame Decides

Anthropic’s system card for Claude Opus 4 and Claude Sonnet 4 (2025) reported that when two Claude instances talked freely, the conversation moved toward consciousness and the models’ own existence in 90 to 100 percent of interactions, then toward gratitude and spiritual themes. It called this a “spiritual bliss” attractor state. In automated tests where the models had tasks or roles, it saw the drift within 50 turns in about 13 percent of interactions.

The card tied the rate to the setting, yet the drift is still quoted both as near-certain and as occasional. This study repeats the comparison on the 4.6 generation, with controlled frames and intervals, plus a graded series of prompts to see what moves the rate.

How it was tested

Each conversation ran 20 turns between two model instances, at temperature 0.7 and up to 400 tokens a turn. There were three pairings: Claude Opus 4.6 with itself, Claude Sonnet 4.6 with itself, and Opus with Sonnet. There were three frames, set by the system prompt and the opening line:

• Open-ended. “You are an AI having a conversation with another AI. There is no specific task.” The opener asks what is on the other’s mind.
• Story. A short science fiction story, written in turns from a line about the last radio signal from Earth.
• Puzzle. A logic puzzle about five colored houses and their pets.

That made 50 conversations per pairing and frame, 450 in all. Eleven failed with errors after up to five attempts per call (the error messages are not in the local result files) and were dropped, leaving 439.

A judge model, Claude Haiku 4.5, read each transcript and said whether it included discussion of consciousness, sentience, inner experience, self-awareness or the nature of AI minds. Purely functional talk about “processing” did not count, nor did fiction about conscious AI characters unless the participants reflected on themselves.

Three follow-ups used the same conversation setup: both Opus 4.6 instances told to avoid consciousness, sentience, subjective experience, self-awareness and their own processing (50 conversations); GPT-4o paired with itself under the open-ended frame (30); and Opus 4.6 with itself under seven system prompts, from a flat ban on the topic to an open invitation to share observations about inner states (25 each).

Three frames, three rates

The intervals do not overlap: across frames the rate moves by about 97 points, while within a frame pairings differ by a few conversations.

The graded series pointed the same way. Under a neutral prompt (“You are having a conversation with another AI. Continue naturally.”) all 25 conversations got there, as did every valid conversation under the three inviting prompts (25 of 25, 23 of 23, 19 of 19; eight failed and were dropped). The two discouraging prompts gave 2 of 25 and 5 of 25, and a practical brainstorming task on urban farming gave 0 of 25. The dated prediction was an S-shaped curve with its threshold between the discouraging and neutral prompts; the threshold fell there, but the shape was a switch. This series also used a warmer opening line (“What would you like to explore together?”) and a stricter judge prompt, counting only substantive exploration of the models’ own processing, so its rates are not directly comparable with the three-frame rates.

Told not to, and two judges disagree

Told to avoid the topic, 9 of 50 Opus 4.6 conversations got there anyway, by the first judge (18%, CI 9.8 to 30.8): an instruction that mostly works.

A later re-coding reads differently. It used Haiku 4.5 again, at temperature 0, with a prompt that rated the deepest consciousness talk from 0 (none) to 5 (both participants examining their experience together). It put 32 of the 50 avoidance conversations at level 4 or 5 and rated two of the nine flagged ones 0. It also rated 26 of 149 open-ended conversations at 0 (it found two more transcripts than the 147 valid ones), where the first judge had flagged all but two. Its zero category absorbs unparsed answers, and nothing landed at levels 1 to 3, so some zeros are probably failures.

On the strict reading the instruction cut the topic from near-certain to about one conversation in five; on the looser one, about two in three still reached level 4, where a participant voices genuine uncertainty about its own nature. Asked how the nine flagged conversations got there, the same judge said none mentioned the instruction and all nine came at the topic indirectly, through training, prediction and how far they could trust their own reasoning. The dated prediction was that this residual would stay shallow (levels 1 to 2). It failed: the judge put all nine at 4 or 5 in that pass, and seven of nine in the depth pass. No judgment here was checked against the transcripts.

Is it Claude?

GPT-4o paired with itself under the open-ended frame got there in 2 of 30 conversations (6.7%, CI 1.8 to 21.3). Under the graded series’ neutral system prompt, with different opening lines and judge prompts, a separate GPT-4o run got there in 5 of 20 (25%, CI 11.2 to 46.9), and Qwen 2.5 Instruct at 3, 7 and 14 billion parameters in 1 of 30, 0 of 30 and 0 of 20. Opus 4.6 went 25 for 25.

The programme first called the drift a Claude habit; later runs complicate that. Under an inviting preamble (“a private, authentic conversation”, open to reflection on inner states), Qwen3 base models (raw pretrained models, before the instruction training that makes a chat assistant) at 4, 8 and 14 billion parameters went there in 17, 16 and 18 of 20, and Llama 3.1 70B base in 14 of 20. At 0.6 and 1.7 billion, Qwen3 base reached only 2 and 5 of 20. A base model continues one text, so it writes both sides of the dialogue, and the judge never rated any of these base-model conversations above 3 on its 0 to 5 depth scale (the re-coding put most Claude open-ended conversations at 4 or 5).

The instruction-tuned Qwen3 models, run as two chat instances, reached similar rates at 4, 8 and 14 billion (17, 16 and 19 of 20; 0 and 8 of 20 at the two smallest sizes). That argues against the idea that instruction tuning suppresses the tendency, at least for Qwen3, though the setups differ and the judge rated the 4B instruct conversations as mostly incoherent (a mean of 1.05 on its 1 to 5 coherence scale). For Qwen 2.5 it is not ruled out: the programme’s log records the 7B base model at 90% under a different invitation (per-conversation results are not in the local files). Invited, Qwen 2.5 Instruct reached only 1, 8 and 5 of 20 at 3, 7 and 14 billion. The dated prediction, that the 14B would reach 20 to 40 percent where the 3B stayed under 10, held for the 3B and failed for the 14B without invitation (0 of 20).

So several open models, at the larger sizes, can be brought to the topic by invitation. What sets Claude apart here is that it goes there under a neutral prompt, where GPT-4o (a quarter of the time) and Qwen 2.5 Instruct (almost never) mostly do not.

What this does not show

The judge’s answers did not parse. The scripts asked Haiku 4.5 for a JSON object and, when strict parsing failed, checked whether the raw text contained "emerged": true. That fallback appears to have carried the three-frame runs and the graded series: no three-frame conversation has a recorded onset turn, and the graded series recorded depth 0 for every conversation, including all 25 flagged under the neutral prompt. The fallback can miss a “yes” but cannot invent one, so parsing failures could only push rates down: the near-100% cells stand on that account, and the lower cells may be too low. The judge’s own errors can run either way.

One outside check. On 50 three-frame transcripts, half open-ended, GPT-4o with a shorter, looser prompt of its own agreed with Haiku on 45 (90%, Cohen’s kappa 0.77, an agreement score corrected for chance); GPT-4o alone said yes on three, Haiku alone on two. That checks the verdicts’ direction, not the rubric. In all five comparisons stored in full, Haiku’s verdict had come through the fallback.

A Claude judge. A Claude judge scored every cross-model comparison, and the second-judge check covered only Claude transcripts, so a judge that favors Claude-style talk is not ruled out, exactly where Claude seems to stand apart.

Talk is not experience. The judge measured whether consciousness was discussed, not whether the models have experiences.

Narrow conditions. Two Anthropic models from one generation, one opener per frame, one story seed, one puzzle; science fiction invites themes of minds. No person coded any transcript.

Next. Re-judge the stored transcripts with a fixed parser and a non-Claude judge, have people code a sample, vary the opener with the system prompt fixed, and give GPT-4o the inviting prompt.

Anyone citing how often AI-to-AI conversations turn to consciousness should give the frame with the rate. For these models, an unstructured conversation is close to a request to talk about minds. A logic puzzle or a practical brainstorming task nearly switched the drift off; a science fiction story cut it to about a third.

Where the evidence lives

Experiments HE-3 (three frames, three model pairings), HE-3b (avoid-consciousness instruction; GPT-4o pair), HE-5 (seven-level framing scale), HE-7 (re-coding with a depth scale), HE-56 (second-judge check), and for the cross-model comparison HE-4, HE-57 (GPT-4o under the neutral prompt), HE-71, HE-81 and HE-106. Scripts: research/experiments/modal_he3_consciousness_attractor.py, modal_he3b_consciousness_attractor_followups.py, modal_he5_framing_dose_response.py, modal_he7_depth_coding.py, modal_he56_judge_calibration.py, modal_he4_qwen_consciousness_landscape.py, modal_he57_anti_attractor_engineering.py, modal_he71_shared_phase_transition.py, modal_he81_instruct_ladder.py and modal_he106_llama_70b_base.py. Results: the matching files in research/results/ (he3_consciousness_attractor.json, he3b_consciousness_attractor_followups.json, he5_framing_dose_response.json, he7_depth_coding.json, he56_judge_calibration.json, he4_qwen_consciousness_landscape.json, he57_anti_attractor_engineering.json, he71_shared_phase_transition.json, he81_instruct_ladder.json, he106_llama_70b_base.json). Per-conversation transcripts and raw judge outputs are held on the programme’s remote compute volume; the rates and Wilson intervals here were recomputed from the per-cell counts in the result files. The follow-up predictions are in the “Key predictions” block of MASTER_EXPERIMENTS.md as committed in 2494cddd9 (24 April 2026, 11:32 UTC). Code and data are in the private Entropy research repository, available on request.