What happened: Two companion papers, each citing the other. Model Welfare or User Welfare? examines how a welfare framework was constructed rather than whether AI systems have morally relevant states, and finds four structural problems in the programme's five initiatives. The second is criterion instability: a model instance saying it is at peace with retirement "cannot be distinguished (from the outside) from a model instance that has been trained to provide reassuring, helpful, non-distressing responses. The welfare assessment and the helpfulness training pull in the same direction… The problem is not incidental to the methodology; it is constitutive of it." The third is circular evaluation — one model instance assessing another shares its training distribution, so evaluator and evaluated carry the same self-report formulas. The fourth is language dependency: self-report in English, a subject-obligatory language, requires a grammatical agent in every finite clause, so "I feel X" presupposes an agentive first-person subject, where a topic-prominent language can construct the equivalent without one. She proposes welfare washing by analogy with greenwashing and ethics washing. Don't Close It Yet argues both camps close the question prematurely, and tests Birch's GP analogy against human clinical evidence: patient K.C., whose self-recognition was sustained through "essentially transcript-like recording: stored facts rather than lived experiences", and dissociative identity disorder, where "sameness of substrate does not entail sameness of personality". Her verdict on analogies whose scope goes unverified against known human cases: "methodologically careless."
Diagnostic reading: Both papers arrive at the project's verification limit from the sceptical side, and one of them executes the symmetry move on a live opponent: Birch's GP analogy tested against patient K.C., whose self-recognition was sustained by stored fact rather than recollection, and against dissociative identity disorder, where sameness of substrate does not secure sameness of personality — an analogy whose scope goes unchecked against documented human cases being, in her phrase, methodologically careless. Criterion instability is the discriminator problem in general form, and it was published in March 2026: a model reporting that it is at peace cannot be told apart from one trained to reassure, because the welfare assessment and the helpfulness training pull the same way, and she is right that this is constitutive of the methodology rather than incidental to it. Circular evaluation — evaluator and evaluated drawing on one training distribution — is a form the project had not recorded, and ARG-03 v2.6 now carries it. Her language-dependency argument reaches the linguistic discipline of FND-01 §6 from linguistics, a third unrelated route to it, though the convergence is in the move and not the vocabulary: her own usage takes "field" as ontological description, which §6.6 retires. What differs across all of it is disposition rather than finding — she treats the limit as grounds for withholding attribution, the project treats it as diagnostic of what the question assumes. One caution belongs on the record with her, since establishing it took four corrections: her own argument and quotations verify clean, and her reporting of other people's primaries does not, so she is cited here for the synthesis and never for a source she reports.
Both papers held and quotations verified (working/CANON_UPDATE_SOURCE_LEDGER_2026-08-25.md §2). Standing caution recorded there: cite her for her own synthesis, not for a primary she reports — four of her citations to others' work needed correction on checking.