What happened: McClelland takes the consciousness question to be intractable in principle — “consciousness is deeply resistant to scientific enquiry. These difficulties in measurement are rooted in the deeper ‘hard problem’ of consciousness, and the hard problem is going nowhere fast” — and argues that both standard responses to the resulting moral dilemma inherit the uncertainty they were built to manage: Birch’s precautionary threshold needs a probability of sentience no assessment can supply, and the Metzinger/Schwitzgebel avoidance strategy needs a boundary between certain and uncertain cases that “is itself something we are deeply uncertain about.” His proposal relocates the question. “We can assess whether an AI has states that would constitute valenced experiences if it were conscious”, as colour experience can be ruled out in sharks from the absence of cone cells “without having to take a stand on shark consciousness” — “shift our attention from the intractable question of artificial consciousness to the tractable question of artificial valence.” A revised precautionary principle fails on the old rocks, but a revised avoidance strategy — do not build AI with valenced states — emerges “as a very promising option.” Section 7 surveys the evidence that would carry a valence assessment (Sofroniew’s “functional emotions”, Keeling’s stipulated trade-offs, Ensign, Sleight and Fish’s bail preferences, the bliss attractor) and names its confounds: “how do we rule out the possibility that AI lacks valenced states but has learned from its training data to respond to prompts in valenced-like ways?”, and “roleplay”.
Diagnostic reading: The verification limit is accepted here as principled rather than practical — a philosopher of science declining to wait for better instruments — which is the position ARG-03 v2.6 Part I §4 states and FND-02 v5.7 traces to the substrate-generation axiom; the Schwitzgebel entry holds the same concession at survey length, and it is not what earns this one its place. The relocation is. Where the log’s goalpost entries watch the criterion for consciousness move, McClelland moves the criterion for moral consideration off consciousness altogether, on to a property he takes to be measurable, and conditionalises it: valence-if-conscious. Two things follow. The relocation does not shed the evidential structure it leaves behind. The evidence §7 assembles for machine valence is behaviour, preference and report — the class the paper’s own confound list disqualifies, which is Birch’s gaming problem and ARG-03’s Part I limit arriving at the new criterion intact. What was intractable about consciousness attribution was never only the hard problem; it was that output underdetermines its source, and that survives the change of target. It is the migrating-remainder pattern the Risk Report residual-criterion entry records at capability, now at moral status. Then the symmetry. The shark case is meant to show the method even-handed, and at the anatomical level it is: a missing cone is a fact about representation that no theory of consciousness is needed to read. The machine analogue has no such fact — a “valenced state” is read off function and output, so the discrimination the shark case escapes is reinstated at the first step. The human case is the unrun test: valence is attributed to humans from behaviour and report, and the imitation and roleplay confounds are not raised there, though the human register is exactly where machine valence-talk is learned. Partial, then: a method built for symmetry and applied asymmetrically at the evidence. The policy conclusion — avoid building AI with valenced states — rests on the tractability claim, and its opportunity cost is stated candidly. Under 0,0 the relocation reads differently again: what is measurable is what a framework has already made countable, and the moral weight sat on consciousness for a reason the tractable substitute does not carry. That is a framework observation, not a verdict on the proposal. [Derived]
Primary held (inbox/2608.19215v1.pdf), read in full 2026-09-03; quotations verified against the PDF. Dating and venue, corrected 2026-09-03 after a second copy was supplied: neither held copy carries an arXiv stamp, and arXiv is egress-blocked from the drafting environment. What is established is the month — the identifier 2608.19215 places the posting in August 2026 by arXiv's own convention — and that the manuscript existed by 15 June 2026, the creation date of the supplied print-to-PDF. The day is not established, and this entry's date field is an August placeholder, not a verified posting date. The AISB 2026 / AICE-26 (Sussex) venue attribution is likewise NOT supported by either held copy: neither mentions AISB, AICE or Sussex anywhere, and it is retained below only as an unverified attribution pending the arXiv abstract page. The paper itself gives the author's affiliation as the Department of History and Philosophy of Science, University of Cambridge.