What happened: Birch sets out a centrist position on AI consciousness: two challenges must be taken seriously at once — that users will misattribute human-like consciousness on the basis of mimicry, and that "profoundly alien forms of consciousness might genuinely be achieved in AI", while our theories remain too immature to say. §4.2 records that he was drawn around 2020 to a parity principle — "any behaviour we take as evidence of consciousness in a living animal, human or otherwise, should also be taken as evidence of consciousness in AI" — and rejects it: "But this principle is false." His reason appeals to no biological premise. It is the gaming problem, which he attributes to Kristin Andrews and himself: chatbots "seek user satisfaction and extended interaction time" and draw on training data to mimic the signs, so they are "incentivized and enabled to game our criteria". Markers so produced "lose their positive evidential value", and "conceivably they might even start to become evidence against consciousness, if the mimicry becomes suspiciously flawless" — by analogy with greenwashing, which he calls consciousness-washing, while noting the disanalogy himself: greenwashing is usually intentional, consciousness-washing "most likely not". He states that the problem "would still arise even for a system that was behaviourally indistinguishable from a real-life human, provided its recreation of human-like behaviour still depended on an extremely rich training corpus". §4.3 turns to theory-driven architectural indicators and meets the Janus problem: "the same evidence can point in opposite directions, depending on one's background beliefs… both responses are reasonable", with a second level — deadlock over whether a finding counts as evidence at all. §4.4, "Escaping two-level deadlock", proposes evaluating biological-naturalist hypotheses by overall fit with the evidence on how consciousness is distributed in the biological world.
Diagnostic reading: Birch names the symmetry move and rejects it, but the principle he rejects is an evidential-transfer rule and the symmetry principle is not one — it tests an objector's criterion rather than carrying evidence from one case to another, a distinction ARG-01 v1.5's scope clause now states in the document rather than leaving to be inferred. What the gaming problem did force is internal: three of ARG-01's applications concluded that an objection discriminates only if biological generation is assumed, while the same document's catalogue of surviving asymmetries said the opposite, and that correction was owed whether or not he had written. The concession is real and now catalogued as criterion-bearing provenance; what it does not carry is the disqualification, which relocates to a linking premise, and the premise generally supplied is kinship — relevant similarity to the assessor, a property of the epistemic relation rather than of the system. The one step the project declines is the sign-flip: for markers to count as evidence against experiencing one needs a calibration of expected marker-display, and the only calibration available is the human distribution, which is the similarity measure the objection set out to replace. The sharpest passage is not the disagreement but §4.4, where the escape from the Janus problem evaluates biological-naturalist hypotheses against the evidence on how consciousness is distributed in the biological world — with the assumption marked in his own footnote, "and I assume here that we can". What underwrites that settlement is kinship again; this is the substrate assumption operating at the level of method, and it is stated by the author rather than imputed to him. On the stance matrix he splits rather than sorts — independently motivated on one marker, late and asymmetric on two others — which is what the markers exist for (FWK-03 v1.5 §1.2.1).
Held at inbox/Birch - AI Centrist Manifesto.pdf, v6 dated 31 July 2026; quotations verified against it (working/CANON_UPDATE_SOURCE_LEDGER_2026-08-25.md §3, §4.2). The paper has revised at least four times since September 2025 — cite by version.