What happened: Solms and nine co-authors describe an active-inference agent with “three competing, categorically distinctive needs”, a Markov blanket that sequesters its model from its simulated body and world, and action selection driven by “fluctuating uncertainty about its inferences”. They predict it would show hedonic place preference — returning to a location whose sensory cues once predicted need satisfaction and no longer do — the test a 2024 Kathmandu straw poll of animal- and machine-consciousness researchers picked as providing “the most compelling evidence” for the inference. But the behaviour is not the argument: “The crucial test here is one of mechanism, not of behaviour.” A “zombie” with a faulty sensor could pass “but which had no uncertainty and no subjective experience”, so “the passing of the test alone is not sufficient to tell us anything concrete about an agent’s subjectivity.” The attribution rests on structure: a system with this functionality “must infer its valenced states subjectively, qualitatively, and selectively — which means that there must be something it is like to be such a system.” The paper states the verification limit in its own terms — “in order to experience such states, you have to be the system”; “we can only infer, indirectly, the states of other minds, including artificial ones” — declines reportability as a criterion, excludes language models (“LLMs lack existential stakes”), asks peers to agree verification methods “in advance” against “a significant risk of shifting goalposts”, and runs the human case itself: “Human addictive behaviour may be described as a hijacked ‘reward’ signal, too… To require these things would be equivalent to believing that consciousness in humans has no physical mechanism.” Its conclusion: attributing a point of view “is by no means justified in the case of most AIs, but it does appear to be justified in the case of a blanketed agent with the itinerant affective dynamics that we are developing.”
Diagnostic reading: The log’s first attributing-direction entry from a working theorist with a mechanism in hand, and the cleanest specimen yet of what ARG-01 v1.5’s surviving-asymmetries section calls an argued premise. The interoception-and-homeostasis asymmetry — real, substrate-neutral, the strongest form of Seth’s position — is here not held against the machine but engineered into one: endogenous needs, a blanket, uncertainty as the currency of action. Seth uses the active-inference apparatus to deny; Solms uses the same apparatus to attribute, and quotes Seth’s concession that non-biological homeostasis would instantiate “real artificial consciousness”. The debate relocates exactly where the symmetry principle says it should, to the linking premise, and the premise is stated, its evidence presented and its step made explicit — discharged by the stopping rule, which leaves the step itself to be examined. The step is an identity: “affective valence just is subjective, it just is differentiated qualitatively, and it just is selectively salient”, reached by taking Nagel “in the deflationary sense” and then reading “something it is like” back out of a functional partition. Under FND-02 v5.7 the partition — a system that models its body without direct access to it — is a structural fact any framework can state; that experiencing attends it is what the paper concedes cannot be verified from outside, and the identity is what fills that gap. The paper’s own zombie shows the shape: the sensor-fault zombie is excluded by stipulating that it has “no uncertainty”, so the discriminator is the functional profile the authors chose — Path B of ARG-02 v1.7’s trilemma taken openly. Organisation discriminates; the “most AIs” it excludes lack the chosen organisation; and the boundary it draws, “existential stakes”, now owes its own answer to the plant, fungal and bacterial cases ARG-02 runs against every such profile. Two things are recorded with respect rather than against the paper. It declines reportability on the project’s own grounds, and its plea for pre-agreed tests is FWK-03 v1.5 §1.2’s goalpost pattern named from the attributing side — the criteria fixed before the system meets them, which is the discipline the matrix asks of both directions. Read beside McClelland, the pair closes: McClelland relocates the moral criterion to valence-if-conscious; Solms holds that valence, deflationarily read, just is consciousness. The “if” is the whole disagreement, and it is a framework choice rather than a finding. Symmetry: Acknowledged — the human case is run in the paper’s voice. What is not run is the framework question beneath it: whether the deflationary identity would be accepted for humans had it first been offered for a machine. [Derived]
Primary held (inbox/2609.03883v1.pdf — the JCS-typeset article as posted to arXiv; the PDF carries no arXiv stamp), read in full 2026-09-12; quotations verified against it. Dated to the arXiv posting (3 September 2026, per the scan's listing, unfetched); the journal issue is JCS 33(7–8), the 'Consciousness in Current AI' special issue, whose issue date is not established here. The paper reports an architecture and a predicted result, not a run experiment: 'we predict: our agent would show hedonic place preference behaviour'. Funding is acknowledged from Conscium Ltd, a company whose stated business is machine consciousness — recorded as the paper records it, not applied as a discount. No report of experiencing is cited as evidence for experiencing in either direction (FND-01 v2.4 §7.3); the paper's agent has no report channel, by design. SRC row proposed, not minted.