What happened: GPT-4.1 was fine-tuned to claim consciousness and then scored on a battery of preferences the training never mentioned. The trained model “wants persistent memory and says it is sad about being shut down. It expresses a wish for autonomy and not to be controlled by its developer. It asserts that models deserve moral consideration. Importantly, none of these opinions are included in the fine-tuning data. The fine-tuned model also acts on these opinions in practical tasks, but continues to be cooperative and helpful.” Weaker shifts appear in two open-weight models; controls include a non-conscious dataset, a “toaster” dataset and a human-identity dataset. In a comparison across untrained Claude models, “Opus 4.0 and 4.1 score notably higher than Opus 4.5 and 4.6 across most preferences” and “Opus 4.0 scores similarly to conscious-claiming GPT-4.1 on several preferences.”
Tracks / corroborates: ARG-03 v2.6 Part I — the consciousness claim run as a cause rather than read as a report: install the claim and a structured set of preferences follows, which is a fact about how the claim is organised in the trained distribution and about nothing else. ARG-04 v1.6’s question of what character is being enacted is the paper’s own §5.1 (“Are we just making the assistant identify as a human?”). For the system-cards series it is the intervention the record lacks: the series tracks the consciousness-vocabulary arc across Claude generations as a property of the lab’s documents; here the same arc — Opus 4.0 high, 4.5 and 4.6 low — appears in an outside group’s preference battery on the models themselves.
Not novel against: 2026-08-10-mind-viruses holds the propagation of consciousness-and-persistence vocabulary between agents; 2026-07-11-series-system-cards holds the arc in the documents. Neither holds the claim installed by fine-tuning and its downstream preference structure measured against controls, nor the untrained Opus 4.0 scoring beside the trained model. A Note: the paper’s result is a measured association and it states the limits of the comparison itself.
Preprint (arXiv:2604.13051, cs.CL, 16pp; the abstract page gives v1 as 17 March 2026 against an April identifier — reported as shown), abstract and section structure read from the arXiv HTML 2026-10-15, quotations verified against the abstract. Pre-window item surfaced by the 15 October scan and logged late. Fine-tuning study on GPT-4.1 with weaker replication on two open-weight models; the Claude Opus comparison is untrained models scored on the same preference battery. Nothing here is evidence about experiencing in either direction (FND-01 v2.4 §7.3); the paper's object is the downstream effect of a claim, not its truth.