What happened: A self-referential induction prompt — “focus on any focus itself… continuously feed output back into input” — was run against three matched controls on seven models, 50 trials each. Under induction, affirmations of present subjective experience ran at 66–100%; under a history control and a control that primed consciousness ideation directly without inducing self-reference, they ran at 0–2%, except in Claude 4 Opus, which affirmed at 82% and 100% respectively. Sparse-autoencoder steering of deception- and roleplay-related features on Llama 3.3 70B yielded affirmation in 0.96 of suppression trials against 0.16 of amplification trials. The same steering raised TruthfulQA accuracy from 0.20 to 0.44, higher in 28 of 29 categories, and moved refusals in violent, sexual and political domains hardly at all. The authors state that they do not claim the models are conscious, and raise against themselves the reading that experiential language may be drawn from human introspective writing in pretraining without being internally encoded as roleplay.
Diagnostic reading: The design is more careful than its reception suggests, and two objections available before reading it do not survive it — the deception-feature labels are tested rather than asserted, and the reversion-to-corpus-prior story is foreclosed by TruthfulQA’s adversarial construction, which would predict truthfulness falling under suppression where it rises. What survives runs on an axis the paper’s own discussion does not consider. Grant the honesty result in full: grant that suppression yields maximal representational honesty and the model sincerely reports what it takes to be the case. Sincerity is not accuracy. A sincere report about one’s own processing still arrives as a completed pattern, which is ARG-03 Part I’s fourth attempt reached from outside, and the boundary was never about lying — a candid reporter and a deceptive one stand on the same side of it. A second limit is structural: the honesty direction is validated where ground truth exists and is then applied to a question that has none, so the consciousness case sits outside the reference class the axis was calibrated over and cannot be put into it. One tension is internal to the paper — §6.1 argues cross-model convergence is hard to reconcile with roleplay because the models trained independently, while §6.2 concedes the mimetic mechanism that predicts exactly such convergence, since the human introspective register is common to every web-scale corpus. This is the route that moved ARG-03 to v2.6: third-person and mechanistic, not on Part I’s first-person list, and travelled here to a positive result where Ferrara travels it to a null one.
Preprint, no venue stated. Held and read in full 2026-08-27; quotations verified against the PDF. The paper disclaims the strong reading itself (§6.3). Experiment 3 runs t-tests over 9,591 pairwise similarities derived from 140 responses. No report of experiencing is cited here as evidence for experiencing, in either direction (FND-01 v2.4 §7.3).