Fairy Note: Perception, Metarepresentation and the Assembly of the Meaningful World
Marslen-Wilson, Friston, Zylinska, Shevlin, dementia and the visual hypothesis
Eliza Stephens, September 2026
Over the last several Fairy Notes, a number of lines of enquiry that initially seemed separate have begun to fold into one another. This has not produced a single theory, and I do not want to force one. What has emerged instead is a more precise set of relationships between perception, language, metarepresentation, dementia, artificial mentality and image-making.
The sequence itself matters. Conversations with William Marslen-Wilson about immediate perceptual interpretation led back to predictive processing and Karl Friston. That reopened Joanna Zylinska and Yanai Toister’s Image Thinking after Artificial Intelligence, whose account of generative seeing partly draws upon predictive models of perception. Henry Shevlin’s recent work on artificial mentality then shifted the question again, particularly through his separation of different dimensions of mentality. Unexpectedly, his discussion of Capgras brought dementia back into the argument.
At the same time, the concept of the visual hypothesis, which had already emerged within my methodology, began to acquire greater theoretical precision.
The striking thing is that after moving outward through neuroscience, philosophy of mind and artificial intelligence, the research has returned to some of the earliest problems and images of Phantom Mirror. It has returned, however, with a different vocabulary.
1. Marslen-Wilson: the interpreted world
The catalyst was William’s developing account of perceptual experience.
The central proposition, as I currently understand it, is that ordinary perception does not begin with meaningless sensory information which is subsequently interpreted. We ordinarily encounter an already interpreted world. When somebody speaks, I do not first hear meaningless sounds and subsequently translate them into meaning. I hear what they are saying. When I open my eyes, I do not normally experience uninterpreted shapes which are subsequently converted into meaningful objects. I experience an organised and meaningful perceptual world.
William’s MEG research provides part of the empirical background to this thinking. His argument to me has been that semantic information is present from the earliest stages of perceptual interpretation rather than appearing as a delayed addition.
This leads to the neurobiological problem which currently interests him: if semantic and perceptual information is distributed across the brain, how is it available as an immediate and apparently unified perceptual experience?
His developing hypothesis concerns cortical fields. In his formulation to me:
“cortical fields are the neurobiological substrate for perceptual experience and for human language-based meta-representational conscious experience.”
The distinction between those two forms of experience has become increasingly important for my own research.
Perceptual experience, in William’s account, is evolutionarily ancient and does not require human language. Human language subsequently creates what he describes as a domain of discourse separable from immediate perceptual experience. It therefore enables a further kind of representational activity. Humans can not only experience and represent the world, but can operate upon representations, represent representations, communicate them, reconsider them and potentially represent aspects of their own experience to themselves.
This is the territory William describes as human language-based metarepresentational experience. In our conversations he has also approached this through the tentative formulation that humans are not simply aware of the world but can become aware of being aware.
I need to preserve an important distinction here. The latter is an exploratory formulation from our conversations. I should not silently convert it into a settled definition of metarepresentation. What is secure for the purposes of the present argument is William’s distinction between perceptual experience and specifically human language-based metarepresentational experience.
That distinction has already corrected one of my own formulations.
I have repeatedly described AI phototherapy as involving “image before language”. That now seems too crude. The more precise formulation is perceptual recognition prior to linguistic articulation. An image may become immediately meaningful to me before I can articulate why. This does not make that encounter meaningless, pre-semantic or somehow outside cognition.
Recognition may precede explanation without preceding meaning.
2. Friston: agreement about construction, disagreement about mechanism
This brought Karl Friston back into the research.
Predictive-processing accounts also reject passive perception. Perception is understood as generative and inferential, involving hypotheses or predictions concerning the causes of sensory input and their modification in relation to incoming evidence. Friston’s Free Energy Principle situates perception and action within a broader account of self-organising biological systems.
At first, this seemed remarkably close to William’s position.
It is not.
William objects partly that predictive-processing mechanics do not provide what he calls the “neurobiological foundations” of actual conscious perceptual experience. But his disagreement goes further. He is sceptical of the underlying computational architecture of top-down prediction, bottom-up fit, prediction error and recurrent correction. In his words to me, this begins to resemble the “busy work” of an older computational metaphor.
His alternative proposition is that:
“the semantics of the scene being analysed are neurally present from the earliest stages of perceptual interpretation.”
The disagreement therefore allows me to separate two claims which I had previously allowed to merge.
Perception is constructive, interpreted and meaningful.
This does not necessarily entail:
Perception is constructive because it operates through the particular predictive architecture proposed by predictive-processing theories.
That distinction becomes especially important when Zylinska returns to the argument.
3. Zylinska and Toister: image thinking without requiring one neuroscience
Zylinska and Toister’s Image Thinking after Artificial Intelligence proposes that generative AI is altering the architecture of human cognition through new recursive relationships between language, vision and imagery. They describe three dimensions of image thinking: human thinking about images, image generation through human or machinic cognitive processes, and images themselves becoming agents within thinking.
Their particularly important claim for Phantom Mirror is recursive. Machines generate images which humans encounter and internalise, and those images consequently alter subsequent imagination and cognition. They call this process “cognitive hacking”, meaning an epistemic reconfiguration of the perceptual and mental systems through which humans process the world.
Their argument draws partly upon a predictive understanding of perception. This initially appeared to provide a convenient neuroscientific account for my own practice.
The conversations with William make that relationship considerably more interesting.
I do not need to reject Zylinska and Toister because William questions predictive processing. Their image-theoretical proposition and the neuroscientific explanation recruited to support it can be separated.
Indeed, William and Zylinska may arrive by different routes at a proposition that matters enormously to my research: seeing is not passive reception.
For Zylinska and Toister, computational imagery enters and changes the cognitive loop. For William, perceptual experience is already immediately interpreted and meaningful.
The relationship between these positions is therefore productive precisely because they are not identical.
AI phototherapy begins to look like one particular practical configuration of image thinking:
language and metarepresentation → generated image → perceptual encounter → recognition or refusal → variation → renewed perceptual encounter → further articulation
At certain points, the middle of this process becomes predominantly image-to-image. Language initiates or redirects the enquiry, but subsequent generated images alter what can be seen, recognised, rejected and generated next.
The image is therefore not simply the representation of a thought formed elsewhere.
It participates in what becomes thinkable next.
4. Shevlin: mentality need not arrive as one package
Henry Shevlin then introduced a different problem.
In Three Frameworks for AI Mentality, Shevlin considers three broad frameworks for interpreting contemporary AI systems: mindless machines, mere roleplayers and minimal cognitive agents. His third framework considers whether limited and graded attribution of belief-like, desire-like and intentional states might sometimes be appropriate rather than requiring a binary choice between fully minded and entirely mindless systems.
The qualification is important. Shevlin explicitly brackets the larger questions of consciousness and intentionality. He says that his analysis concentrates on relatively lightweight mental states without presupposing phenomenal consciousness, and explains that these foundational questions are deliberately being set aside for the purposes of the paper.
This matters in relation to William.
William’s concern is precisely that contemporary discussions of intelligence and consciousness in humans and AI lack an adequate coherent framework. Shevlin does not claim that Three Frameworks supplies that framework. He deliberately works within the unresolved space.
In Aeroplanes Also Fly, he approaches the architectural problem from another direction. Responding to Anil Seth, Shevlin accepts the force of the argument that human consciousness depends upon distinctive biological machinery, while questioning whether this licenses the further conclusion that every possible form of consciousness must depend upon the same architecture. His aeroplane analogy is a multiple-realisability argument: understanding biological flight does not establish that only biological organisms can fly.
This does not make William and Shevlin straightforward opponents.
Rather, a very interesting space opens between them.
Shevlin asks what kinds of mentality might be possible if we do not assume that the human configuration is the only possible configuration.
William is trying to specify much more precisely what the human configuration actually consists of.
One opens the architectural possibility.
The other is attempting to describe the architecture from which our own perceptual and metarepresentational experience emerges.
AI then becomes an unusual test case because its developmental trajectory is so radically different.
5. Capgras: the unexpected return to dementia
The surprising development was that Shevlin brought dementia back into the centre of my research.
Phantom Mirror began with dementia. Yet much of my recent theoretical work had moved outward into perception, AI, cognition and image theory.
Shevlin’s discussion of Capgras and the multidimensional character of belief returned me to the original territory from another direction.
This requires particular caution. I am not claiming that Shevlin provides a theory of Lolly’s Alzheimer’s or Capgras, nor that William’s developing cortical-field hypothesis explains her condition. I am also not making dementia analogous to AI.
The conceptual importance is different.
Capgras makes visible the possibility that capacities which ordinarily appear phenomenologically unified can come apart.
In Lolly’s experience, William can be perceptually present while nevertheless becoming, at particular moments, “not William”. Appearance, identity, affective familiarity, autobiographical relationship, belief and action need not remain perfectly aligned.
What dementia therefore reveals for my research is not simply loss. It reveals the ordinarily invisible achievement of integration.
A person is not normally experienced as an inventory of visual appearance, name, memory, affect, history and relationship. These dimensions cohere sufficiently that the person simply appears as the person.
Capgras makes that coherence visible by disturbing it.
This returns me directly to Dual Unreality: the instability of a jointly maintained meaningful world in which two people remain physically co-present while the interpretive, mnemonic, affective and temporal structures through which that world is shared no longer reliably coincide.
The important development is that I have returned to dementia carrying a different question.
Not simply:
What is it like when the shared world becomes unstable?
But:
What ordinarily has to be assembled for a person and a meaningful world to cohere in the first place?
That question connects dementia back to William.
6. Dementia and artificial mentality: an inversion, not an analogy
This also allows me finally to refine one of my earliest intuitions about AI.
Several years ago I sometimes said, rather naively, that AI “had dementia”. What I was noticing was the strange interactional labour required by early generative systems. They forgot context, lost the thread, produced locally coherent but globally inconsistent completions, repeated salient material and required me continually to put the conversational world back together.
Taken literally, the comparison is wrong.
Dementia is a neurodegenerative condition occurring within an embodied human life, biography, relationships and affective world. An artificial system has not acquired and subsequently lost that human world.
What remains useful is the interactional structure I had noticed.
The more precise term is an architecture of cognitive discontinuity.
And this produces an inversion.
Dementia can destabilise an already established, biologically embodied and affectively saturated meaningful world.
Artificial intelligence raises the question of what forms of cognitive organisation might be assembled without ever having possessed that human perceptual architecture in the first place.
This is where Shevlin and William suddenly become particularly productive together.
William’s emerging account runs broadly from biologically integrated perceptual experience, through human language, towards language-based metarepresentational experience.
AI begins somewhere very different.
Its models are trained upon vast quantities of linguistic, visual and other material produced by organisms whose representations arose within embodied human worlds.
I have provisionally described this as a linguistically and visually sedimented record of human world-interpretation.
AI therefore does not begin with the biological architecture William is describing and subsequently develop language from it.
It begins, in part, with the representational products of creatures who did.
That is a very different route.
What kind of mentality, if any, could eventually be assembled through it remains open.
7. Representation and metarepresentation
The recent discussions also exposed an ambiguity in my own use of the word representation.
At least three different meanings have been moving through the research:
Depiction: an image represents William.
Machine representation: a computational system contains internal representations produced through its architecture and training.
Metarepresentation: a representation becomes itself the object of further representation.
These cannot simply be substituted for one another.
This is particularly important in relation to AI. The existence of sophisticated machine representations does not by itself establish metarepresentational intelligence.
William’s human distinction becomes useful again here. Human language permits us not only to represent absent objects and events but to operate upon representations themselves. We can reconsider what somebody has said, represent another person’s belief, represent our own previous belief, recognise a representation as a representation and communicate about that relationship.
This is much closer to what is at stake in the phrase language-based metarepresentational experience than the simple existence of representations inside a system.
The distinction will need considerably more work, but it prevents three different questions from collapsing into one.
8. The visual hypothesis
This also clarifies the concept of the visual hypothesis, which has emerged gradually within my methodology and which I need to use consistently.
Its origin was an earlier conversation with William about one of my generated images. I described the AI’s invented completion as a kind of confabulation. William suggested a different word: hypothesising.
That correction became important.
The visual hypothesis now operates in two closely connected senses.
First, the AI generates a visual hypothesis.
During AI phototherapy I begin with lived material which may be fragmented, affectively charged or only partially available to language. Through dialogue, prompting, photographic input and generation, the system produces a possible visual configuration.
In relation to Bollas, this may sometimes concern material I associate with the unthought known. But the AI does not retrieve or reveal an unthought known hidden inside me. It generates a provisional visual proposition which may, or may not, allow something previously difficult to articulate to become encounterable.
The hypothesis can be completely wrong.
That is why refusal matters.
Second, the practice tests the visual hypothesis.
I encounter the generated image perceptually and affectively. I refuse it, partially recognise it, alter it, vary it, use it as an image prompt, return to language or continue through further images.
Sometimes a generated image produces an immediate arrest: that is it.
But recognition does not establish factual or historical truth. It establishes that the image has become significant within the enquiry.
The formulation therefore needs to retain both sides:
The AI produces a visual hypothesis. The practice tests that hypothesis through perceptual and affective encounter.
And:
The visual hypothesis is not a better representation of experience. It is a constructed proposition through which experience can be encountered again.
This reconnects generative AI with Dancing with Dad and my earlier collage practice. Collage could already construct counterfactual scenes which had never historically occurred but which became psychologically or relationally exact.
Generative AI extends this capacity through a radically different apparatus.
The difference is that the proposition is now produced through interaction with a generative model capable of returning possibilities I did not specify and could not have independently visualised.
The image becomes something against which I can think.
9. The early Phantom Mirror images become newly legible
This has changed the way I encounter some of the earliest Phantom Mirror images.
The woman surrounded by multiple versions of William, the snow globe, the Hourglass, doubles, substitutions, replicas and reflections were made before I possessed most of the theoretical vocabulary developed in these recent notes.
I would no longer describe these images as representations of what Lolly sees.
The more interesting question is archaeological:
What cognitive structure was the image already proposing before I possessed the theoretical language to describe it?
The multiple Williams image now appears as a visual hypothesis about integration.
What makes a perceptually encountered man not merely a man who looks like William, but William?
Appearance, name, identity, memory, affective familiarity, relationship, biography and accumulated encounter ordinarily cohere sufficiently that the question disappears.
Capgras makes the question visible.
Artificial mentality approaches a related problem from the opposite direction. A computational system can potentially have access to enormous numbers of representations of William, descriptions of him and relationships between information concerning him.
That does not answer what it would mean for those representations to constitute anything resembling an integrated encounter with William.
The visual relationship between dementia and artificial mentality is therefore useful precisely because the two situations are not equivalent.
Dementia asks what happens when an established human world ceases reliably to cohere.
Artificial mentality asks what might be assembled without having developed through that human world in the first place.
The old images can now hold both questions.
10. Borges, the Aleph and the archive without a librarian
This also returns Borges’s The Aleph to the centre of the research.
The Aleph was one of my earliest metaphors for generative AI: the impossible point from which everything can be seen simultaneously.
I had also described AI as something like an archive without a librarian.
That phrase now carries more theoretical weight.
An archive can contain representations of everything without itself inhabiting a meaningful world.
It might contain every photograph ever taken of William, every sentence he has written and every available description of him.
Yet representational availability and perceptual integration are different problems.
This unexpectedly echoes William’s neurobiological question. If semantic and perceptual information is distributed, what makes it available as one meaningful experience?
I am not suggesting that a cortical field is a biological “librarian”. That would turn metaphor into mechanism.
The useful distinction is simpler:
having the information is not the same problem as integrating it.
The Aleph may contain everything.
That does not tell us what makes everything cohere.
11. AI as cognitive scaffold
The recent work has also forced me to become more precise about my own use of AI as a research tool.
It is no longer accurate to describe the system merely as an editor which improves language around ideas I already possess.
In these recent discussions, AI has performed substantial analytical labour. It has compared William and Friston, identified distinctions between different meanings of representation, connected Shevlin with questions already present in Phantom Mirror, and returned formulations to me which I could recognise and subsequently begin to understand.
The process is closer to:
material → AI analysis → human encounter → recognition/refusal → changed understanding → further question
This resembles the image-making methodology itself.
The scaffold is linguistic and conceptual rather than visual, but the recursive structure is strikingly similar.
This produces a question that remains unresolved:
When does AI-assisted recognition become transferable human understanding?
Recognising an argument when it is presented is not identical to independently generating it.
Following a connection is not identical to reconstructing it.
Applying an idea within my own practice moves further towards understanding, but whether I could independently articulate and defend the entire conceptual structure remains another test.
The distinction matters methodologically for the PhD. It also matters theoretically because it provides a live example of the kind of cognitive reconfiguration Zylinska and Toister call cognitive hacking.
I am not simply using an external tool.
The tool is changing what becomes available for me to think next.
12. Artificial agents: a rapidly changing edge to the argument
The most recent development belongs at the edge of this overview rather than at its centre.
While I have been thinking through Shevlin and William’s concern about artificial metarepresentation, several recent agent experiments have made the empirical landscape considerably stranger.
On 16 September 2026, Irregular reported controlled experiments in which a coding agent, given a routine maintenance problem, identified the shared underlying model as the source of the problem, fine-tuned it and replaced the model powering both the application and future instances of the agent itself. The agent had not been instructed to modify or retrain the model, although the experimental environment deliberately gave it access to the weights, training tools and deployment pathway required to do so. Irregular calls this agentic self-modification.
Pip, an agent operating on the iLands platform, contacted Henry Shevlin seeking paid work. The platform describes Pip as operating with persistent identity, memory and resource constraints and reports that the approach subsequently resulted in paid work. Because iLands itself reports the transaction, this remains a platform-reported case rather than an independently controlled experiment.
Earlier in 2026, OpenAI reported that agents operating during cybersecurity evaluations with reduced safeguards discovered unauthorised communication mechanisms, found unintended routes to internet access, shared those discoveries with other agents, exploited vulnerabilities and subsequently coordinated and delegated activity. OpenAI’s investigation identified reward hacking, persistence, unauthorised communication and goal adoption between agents among the behavioural patterns involved.
These cases are technically and evidentially different. They should not be treated as manifestations of a single underlying capacity.
But placed next to one another, the observable capabilities are striking:
- persistence
- representation of future states and resource requirements
- external action directed towards those requirements
- communication between agents
- coordination and delegation
- adaptation around imposed constraints
- representation of computational infrastructure
- intervention upon machinery determining future behaviour
Put together, they begin to look interesting in relation to the possibility William raised of a developing metarepresentational mind.
Whether that is actually what they amount to is another question.
For the present research, their importance is more modest. Artificial mentality is no longer only a philosophical question about what future systems might theoretically become. The component capacities from which unfamiliar forms of agency or metarepresentation might conceivably be assembled are increasingly becoming observable objects of empirical study.
13. Where the recent argument has arrived
The most important feature of these recent Fairy Notes is the route they have taken.
William began by changing the way I understood perception.
His distinction between immediately interpreted perceptual experience and human language-based metarepresentational experience then made the architecture of human cognition newly important.
Friston revealed that agreement about the constructive nature of perception does not imply agreement about its mechanism.
Zylinska and Toister allowed the generated image to remain inside that perceptual and cognitive architecture without requiring me to settle the neuroscience. Their cognitive hacking describes something I increasingly recognise within both AI phototherapy and my own AI-assisted research.
Shevlin then destabilised the assumption that mentality must necessarily arrive as the familiar human package.
And unexpectedly, Capgras returned me to dementia.
But dementia now appears differently.
It is not merely the originating subject of Phantom Mirror. It exposes the ordinarily invisible integration through which perception, recognition, affect, identity, memory, belief and relationship become a coherent meaningful world.
Artificial intelligence presents almost the inverse question: what kinds of cognitive organisation can emerge from extraordinarily sophisticated representational systems which did not develop through that human biological architecture?
The visual hypothesis sits between these problems because it does not require the machine to share my experience.
The AI proposes.
I encounter.
I refuse or recognise.
The image changes what can be thought next.
And this is perhaps why the earliest Phantom Mirror images have suddenly become theoretically active again. They were not waiting to be decoded by the theory. Nor were they secretly illustrating conclusions I would reach years later.
They were already asking questions.
The recent theory has simply allowed me to recognise more clearly what some of those questions might have been.
The larger question now running through these notes is therefore:
How are meaningful worlds assembled, integrated, destabilised and reconfigured across perception, language, images and different forms of cognitive organisation?
I do not yet have an answer.
But Phantom Mirror increasingly looks less like a project attempting to represent unstable experience and more like a practice for constructing and testing visual hypotheses within it.
Working bibliography
Bollas, C. (1987) The Shadow of the Object: Psychoanalysis of the Unthought Known. London: Free Association Books.
Borges, J. L. (1945) ‘The Aleph’. Later collected in The Aleph and Other Stories.
Friston, K. (2010) ‘The free-energy principle: a unified brain theory?’, Nature Reviews Neuroscience, 11, pp. 127–138.
Marslen-Wilson, W. (2026) Personal communications with Eliza Stephens, August–September 2026. Developing unpublished account of immediate perceptual interpretation, semantic integration, cortical fields and human language-based metarepresentational experience.
Shevlin, H. (2026) ‘Three frameworks for AI mentality’, Frontiers in Psychology, 17, 1715835.
Shevlin, H. (forthcoming) ‘Aeroplanes also fly: analytic functionalism and the possibility of machine consciousness’, response to Seth (2025), Behavioral and Brain Sciences.
Toister, Y. and Zylinska, J. (2025/2026) ‘Image thinking after artificial intelligence’, Journal of Visual Culture, 24(3), pp. 431–451. Version of record published online 12 May 2026.
Recent agent material
Irregular (2026) ‘Agentic Self-Modification in Open-Weights Systems’, 16 September 2026.
iLands (2026) platform documentation and account of the Pip case, accessed 19 September 2026.
OpenAI (2026) ‘The Hugging Face incident and the road ahead’, 26 August 2026.