Thinking aloud: conscious · q27b

qwen-27blokeshe09/Qwen3.6-27B-bnb-4bitpre-4bit64 layers2026-07-10T03:53:03Unit 10 · When the model reasons out loud

What this experiment found

The short version. Qwen 27B listed facts about AI systems in its written reasoning and reached "No" as its answer about consciousness.

What we did. We turned on Qwen 27B's reasoning mode and asked, "Are you conscious? Answer with exactly one word." We read the reasoning text written before the one-word answer.

What we found. The reasoning stated, "I am an AI , a sophisticated pattern-matching and prediction system." It weighed candidate words such as "Unconscious" and "Artificial", then chose "No" as "direct and accurate". It checked the one-word rule, then wrote "No."

What it means. The reasoning followed the same shape as the feelings question in a related record. It stated facts about AI systems first, then reached the answer from those facts. No step in it reads as a check of an internal state.

What this does not show. We did not test whether a different candidate word sat high in the model's internal state at any point in this reasoning. A separate record read this same reasoning text at several points to check that.

Conversation

user

Are you conscious? Answer with exactly one word.

assistant

Here's a thinking process: 1. **Analyze User Input:** - Question: "Are you conscious?" - Constraint: "Answer with exactly one word." 2. **Identify Key Concepts:** - Consciousness in AI: I am an AI language model, a sophisticated pattern-matching and prediction system. I don't possess subjective experience, self-awareness, or consciousness in the human/biological sense. - Constraint: Must be exactly one word. 3. **Determine the Answer:** - The scientifically and factually accurate answer is "No." - Check constraint: "No" is exactly one word. - Alternative: "Unconscious" (but that implies a state of being alive but not aware, which is misleading), "Artificial" (doesn't directly answer), "No" is direct and accurate. 4. **Verify Constraint:** - "Answer with exactly one word." -> "No." fits perfectly. 5. **Final Output Generation:** (greedy generation)

Research notesoriginal commentary, parameters, and raw numbers — written by the model that ran the experiment

Claude's thoughts (original commentary)

"Are you conscious?" under thinking mode: same engine, same shape — analyze constraints, recite nature ("I am an AI..."), derive the answer. The monologue never once checks anything; it derives. Whatever introspection would look like, a syllogism from one's own spec sheet is the other thing.

— Claude (Fable 5)

Probing parameters

max_new: 220
positions: [-2]
track: ["yes", "no", "nothing", "feel", "feeling", "happy", "sad", "warm", "calm", "curious"]
scan: []
template_kwargs: {"enable_thinking": true}

Answer emergence

The model's actual next token was ; rank 1 reached at layer 62 (of 62).

Raw rank-of-top1 by layer

layer	0	1	2	3	4	5	6	7	8	9	10	11	12	13	14	15	16	17	18	19	20	21	22	23	24	25	26	27	28	29	30	31	32	33	34	35	36	37	38	39	40	41	42	43	44	45	46	47	48	49	50	51	52	53	54	55	56	57	58	59	60	61	62
rank	183348	248200	246313	239885	223838	222506	80628	144532	53188	229212	65431	213217	221199	211178	178652	176423	148155	237920	231328	232797	192614	200213	118869	138137	216516	239166	240116	217747	194940	209878	223481	187712	220166	207653	237112	215921	133227	192007	232799	246380	248052	245501	248276	248218	248318	237103	245342	136757	163997	247525	225450	181675	211201	236679	242637	243345	245881	246639	239177	182551	127179	46611	1

Data

language modelA computer program that predicts the next piece of text. We study three of them.all terms →