On 3 September 2026, Mark Solms and nine co-authors submitted an open version of “Inferring Affective Consciousness in an Artificial Agent: A Case Study” to arXiv. The article had already appeared in the 2026 volume 33, issues 7–8, of the Journal of Consciousness Studies; the TU Delft Research Portal classifies it as a peer-reviewed journal article. This is not a study of a large language model describing its own mental state. It examines a comparatively simple artificial agent with three competing endogenous needs. The agent can infer the condition of its simulated body and environment only through restricted sensory inputs, and it adjusts its actions using uncertainty, expected outcomes and prior experience.
The authors are concerned with affective consciousness, not merely whether a system can produce the words “I am in pain”. They begin from hedonic place preference in animal research: an animal may repeatedly return to a location associated with a pleasurable stimulus even when that stimulus has no nutritional value. The paper proposes that an artificial agent could develop an analogous preference when sensory cues mislead its assessment of its needs. However, the paper mainly presents the mechanism and predicted outcome of this minimal test. It does not report a completed set of artificial-agent experiments available for independent replication. The authors also state that passing the place-preference test alone would not establish subjectivity; a faulty sensor could produce the same outward behaviour.
The real question is therefore not whether a machine can act as though it feels. It is whether an artificial system should be judged capable of feeling once it has endogenous needs, treats its own continued operation as an action condition, lacks direct access to its true bodily state, and continually selects and revises policies through uncertainty.
Two meanings of “subjective” must be separated. The first is a functional internal perspective. The agent receives only the information supplied through its sensory interface, and its internal model can diverge from the actual state of its body and environment. The same cue can therefore have one significance to an external observer and another causal role in the agent’s own action selection. The system occupies an informational position relative to its own model. The second meaning is phenomenal subjectivity: there is something it is like for the system to undergo the state. The first can be investigated through architecture, state updates and causal intervention. The second concerns experience itself. Limited access, mistaken inference and self-maintenance create an internal perspective, but they do not conceptually close the distance between these two forms of subjectivity.
Solms and colleagues offer a more disciplined argument than treating coherent language or self-report as evidence of consciousness. They ask investigators to examine the mechanism that produces behaviour. Are needs endogenous variables that the system continuously regulates? Does the sensory interface genuinely restrict its access to its own state? Does uncertainty affect need prioritisation and action selection? Does past experience modify subsequent policies? The paper also recognises that test results can have competing explanations, so success on one behavioural test is not decisive. This moves the discussion from surface resemblance towards inspectable causal organisation.
The further step, however, is still contestable. The authors interpret need regulation under restricted information as a functional basis of feeling, and that inference relies on a disputed functionalist premise. A variable can be indispensable to control without being experienced. A thermostat also changes its actions in response to deviation, while a more complex agent can preserve several needs, rank them and learn. Greater complexity can improve adaptive behaviour, but an accumulation of functions does not by itself explain why phenomenal feeling exists. The opposite requirement—that an artificial system reproduce all biological organisation—would also be too strong, because non-biological realisation cannot be ruled out in advance. The defensible position is neither immediate affirmation nor immediate denial. Evidence must discriminate a consciousness explanation from functionally equivalent non-conscious alternatives.
One established approach in AI-consciousness research is to derive indicators from theories: identify testable architectural or computational conditions implied by several theories of consciousness, then assess which indicators a particular system satisfies. The peer-reviewed work by Patrick Butlin and colleagues uses these indicators to adjust credence in consciousness, not as a binary proof based on any single condition. The Stanford Encyclopedia of Philosophy likewise distinguishes phenomenal consciousness from access consciousness. Information being broadly available for a system’s control does not, by itself, show that the system has qualitative experience or that there is something it is like to be in that state. Both sources underline the need for bridging evidence between mechanistic relevance and a phenomenal conclusion.
In Sustenesis Theory, the study adds a formation structure that can be analysed. Difference first appears in the distinguishability between the system’s internal model, its simulated body and the external environment. Without that difference, the agent would not infer from its own position and could not be mistaken. Constraint does not mean limitation in the ordinary sense. It refers to the specific conditions that make some modes of formation possible and others impossible: sensory bandwidth, a Markov-blanket-style interface, endogenous preferences, resource changes and action channels jointly determine what the agent can know and how it can choose. Sustained Coherence concerns whether the system can preserve relations among its needs, correct its model and remain effectively organised as needs deviate, feedback arrives and policies fail.
These conditions support a functional self-relation. The system’s variables are not arbitrary numbers; they continually affect how it distinguishes favourable from unfavourable conditions, uses experience and revises action. They are closer to the causal evidence consciousness research needs than a one-off linguistic declaration. Yet Sustained Coherence cannot be rewritten as “the structure persists sufficiently well, therefore the system must feel”. Sustained Coherence establishes that an organisation is maintained through operation. Whether phenomenal feeling is part of that organisation must still be assessed through evidence that excludes alternative mechanisms.
My judgement is that endogenous needs, restricted self-access and uncertainty regulation make this artificial agent a serious candidate for testing affective consciousness, but the published evidence does not establish that it already feels. The paper supplies a stronger research object than surface behaviour and identifies mechanisms open to intervention. It does not complete the inference from a functional internal perspective to phenomenal subjectivity. That distinction is especially important because the central behaviour is still presented mainly as a prediction and mechanistic analysis rather than as an openly replicated experimental result. Describing a mechanism that might constitute feeling as evidence that the system already feels would exceed the evidence.
A stronger assessment would require several converging kinds of material: completed and openly reported preregistered tests; interventions that remove or alter need variables, sensory partitioning and uncertainty updates; evidence that the relevant internal states continue to shape choices in novel environments rather than merely track designer-specified reward proxies; control systems that reproduce the same behaviour without the proposed mechanism; and checks for both support and counterevidence from indicators derived from different theories of consciousness. These measures would still not read feeling directly as a thermometer reads temperature, but they would narrow the inferential gap between having an internal perspective and having phenomenal experience.
This conclusion is limited to the architecture and public evidence described in the paper. It neither treats the self-narratives of present large language models as proof of consciousness nor claims that a non-biological system could never feel. The research advances the discussion by replacing personified language with intervention-ready causal comparisons. Its boundary is equally clear: a candidate mechanism is not yet a phenomenal fact.
References
https://arxiv.org/abs/2609.03883
https://doi.org/10.53765/20512201.33.7.014
https://www.cell.com/trends/cognitive-sciences/fulltext/S1364-6613%2825%2900286-4
https://plato.stanford.edu/entries/consciousness/
Discover more from Geoffrey Chen
Subscribe to get the latest posts sent to your email.