AI Dementia and the
Metarepresentational Mind

Cognitive discontinuity and the labour of holding a shared world together

Fairy Note   Eliza Stephens   18 September 2026

This note begins with something I used to say about artificial intelligence that sounded naive even to me. I would say that AI had dementia. I did not mean that a machine literally had a neurodegenerative illness, nor that its condition was ethically or experientially equivalent to that of a person living with dementia. I was trying to describe the strange relational experience of working with early generative systems. They forgot. They lost the thread. They produced locally convincing statements while failing to maintain the wider world in which those statements belonged. I had to keep putting the world back.

Two and a half years later, conversations with my father, Professor William Marslen-Wilson, and my reading of the philosopher Henry Shevlin have given me a more exact way to revisit that early intuition. The useful claim is not that AI has dementia. It is that early generative AI produced a dementia-like interactional structure, marked by discontinuity, weak integration and the repeated labour of contextual repair. This comparison becomes particularly revealing when placed beside Marslen-Wilson’s account of human language-based metarepresentational conscious experience and Shevlin’s argument that artificial systems may possess limited or unfamiliar forms of mentality.

The early intuition

When I first began working intensively with chatbots and Midjourney, the systems had no continuous memory between conversations. A new chat did not continue the previous world. It began without the people, relationships, events and meanings that had already been established. Even within one conversation, the context could deteriorate. Details became confused, earlier decisions disappeared, and invented connections entered the exchange. Eventually I would abandon the chat and begin again, knowing that beginning again meant telling the whole story again.

Midjourney created a related difficulty. I often repeated the same word or phrase because repetition seemed to give it greater hierarchical importance within the image. Technically, this was closer to reinforcing salience or weighting an instruction than compensating for memory loss. Experientially, however, it formed part of the same pattern. If I did not keep restating what mattered, it slipped from the developing visual field.

I sometimes called AI demented and sometimes neurodivergent. These were imprecise terms, and both risked turning lived human conditions into metaphors for technical irregularity. Yet the observation beneath them was not frivolous. I was noticing that this intelligence did not cohere in the way I expected an intelligence to cohere. It could be dazzling in one moment and apparently unable to retain the conditions of that brilliance in the next.

Maintaining the conversational world

The most important similarity was relational. Talking to early AI could feel like talking to somebody with dementia because I was repeatedly required to restore orientation. I supplied names, histories and relationships. I corrected substitutions. I reminded the system what had happened and why it mattered. I became responsible for continuity.

This resembles part of what I do with Lolly, although the ethical and experiential realities are profoundly different. With Lolly, continuity belongs to a living person whose established world is becoming unstable. The history remains real even where access to it is impaired. Her relationships, fears and attachments remain affectively charged. When she asks the same question or says that William is not the real William, the repetition belongs to an embodied life and to a relationship in which both people can suffer.

The early chatbot had not lost an autobiographical world. It had never possessed one. Its discontinuity was architectural rather than neurodegenerative. It did not grieve the lost context or experience anxiety when the conversation fractured. What resembled dementia was therefore not an inner disease state but the structure of the interaction and the work demanded from the other participant.

AI and dementia can both reveal how much work another person must perform to maintain a shared world.

This is the part of my early intuition that survives critical examination. The comparison is not clinical or ontological. It does not say that a machine and a person with dementia are the same kind of being. It identifies a limited functional and relational resemblance involving continuity, contextual retention, integration, confabulatory completion and the labour of repair.

Perceptual experience and metarepresentational consciousness

Marslen-Wilson’s developing account distinguishes immediate perceptual experience from a specifically human language-based form of conscious reflection. His formulation is that cortical fields provide the neurobiological substrate both for perceptual experience and for human language-based metarepresentational conscious experience.

Cortical fields are the neurobiological substrate for perceptual experience and for human language-based metarepresentational conscious experience.
William Marslen-Wilson, personal communication, 1 September 2026

Perceptual experience, in this account, does not begin as a meaningless sensory display awaiting interpretation. A person opens their eyes and encounters an already interpreted world. When somebody speaks, the listener does not first hear meaningless sound and then attach meaning to it. As my father puts it, when he opens his mouth and speaks, I hear meaning.

Human language then permits another level. It allows an experience, belief, memory or interpretation to become the object of further thought. I can see my father, but I can also think about my perception of him. I can wonder whether memory is altering what I see. I can represent what somebody else believes, compare their belief with mine, consider whether either of us is mistaken and describe the difference to a third person. This is metarepresentation: not merely having a representation, but representing a representation as a representation.

This language-based domain is partly separable from immediate perceptual experience. It enables reflection, narrative, counterfactual thought and the representation of other minds. It does not replace the perceptual world from which human experience arises, but it allows us to step back from that world and think about how it is being experienced and interpreted.

The warning

On the morning of 18 September 2026, my father told me that he was now worried that AI might develop a metarepresentational mind. He said, with deliberate force, “If they do get a metarepresentational mind, then we’re doomed.” The warning was not simply that AI is trained on human representations. A text model already operates upon the external products of human linguistic and cultural activity. That fact alone does not give it a metarepresentational mind.

The more serious possibility is that a system might begin to treat representations reflexively. It might model its own informational states, distinguish them from the states of another agent, represent what a person believes, recognise a discrepancy between perspectives and calculate how its communication could alter another mind. That would be more than the production of plausible language. It would resemble the higher-order organisation that language enables within human cognition.

My father’s concern is especially sharp because the developmental order would be reversed. Human metarepresentation develops within an embodied organism already inhabiting a meaningful perceptual world. An artificial system begins with the recorded products of human interpretation. It might conceivably develop a higher-order, language-based organisation without the biological perception, vulnerability, care and world-maintenance from which the human capacity evolved.

This remains a possibility rather than an established account of present systems. Neither sophisticated self-description nor successful modelling of another person’s beliefs proves subjective experience. My father has also repeatedly stressed how poorly consciousness and intelligence are understood. His warning identifies a threshold that concerns him; it does not establish that the threshold has already been crossed.

Henry Shevlin and another way of having a mind

Henry Shevlin approaches the question from philosophy rather than neuroscience. In Aeroplanes Also Fly he accepts that human consciousness depends upon the distinctive machinery of biological life, but argues that this cannot settle whether a radically different artificial system might realise some exotic form of consciousness. Birds and aeroplanes fly through different mechanisms. The existence of biological flight does not make mechanical flight conceptually impossible.

In Three Frameworks for AI Mentality, Shevlin examines three positions. The first treats LLMs as mindless machines. The second treats their apparent mental states as mere roleplay that can be useful without being literally true. The third treats them as possible minimal cognitive agents to which limited and graded beliefs, desires or intentions might sometimes be appropriately attributed. Shevlin does not declare present LLMs fully minded or conscious. He argues that the first two positions are less secure than they initially appear and that mentality may not be an all-or-nothing human package.

His alien example makes the point especially clearly. Imagine an advanced civilisation with art, culture and complex social relationships, but with biology and psychology radically different from ours. Its members might lack human hunger, pain, emotion and perception. We would need new psychological concepts for their unfamiliar forms of life. Nevertheless, it would be extremely difficult to recognise them as intelligent agents without attributing some coarse-grained beliefs, desires and intentions to them.

This does not mean that Shevlin and Marslen-Wilson make the same claim. Shevlin opens the philosophical possibility of nonhuman mentality and argues for graded, multidimensional interpretation. Marslen-Wilson identifies a more specific higher-order possibility, the metarepresentational mind, and responds to it as a danger. Shevlin opens the door. My father worries about what could cross the threshold.

Capgras and the dimensions of belief

The part of Shevlin’s paper that most startled me was his use of Capgras syndrome. Lolly lives with Capgras. She can look at William and insist that he is not the real William, but an impostor or duplicate. This is not merely an incorrect proposition. It reveals a disturbance in the relations among perception, affective familiarity, identity, belief and the wider construction of reality.

Shevlin uses Capgras to challenge the idea that belief is a single, indivisible mental category. A Capgras conviction can be strongly asserted while remaining incompletely connected to the person’s wider account of how the world works. It may be resistant to evidence yet inconsistently translated into other inferences and actions. Philosophers have therefore described some delusions as halfway states between paradigmatic belief and imagination.

On this account, beliefs and belief-like states can vary across several dimensions:

• persistence over time
• responsiveness to evidence
• integration with other beliefs
• capacity to guide action
• inferential consequences
• affective significance

This is where the comparison with artificial mentality becomes intellectually useful. An LLM state may be belief-like along some dimensions and unlike human belief along others. It may support coherent inference within a context window while lacking long-term persistence. It may alter rapidly following new input. It may guide an action without carrying the affective investment that makes the revision of a human belief painful. Shevlin’s framework permits comparison without equivalence.

Capgras therefore joins the two halves of my research without making dementia and AI the same. Lolly’s experience exposes fractures within human belief and recognition. Artificial intelligence asks whether belief-like organisation might appear in a system whose architecture was never human in the first place. Both unsettle the assumption that mind is a single, unified possession, but they do so for entirely different reasons.

From the demented machine to cognitive discontinuity

I would no longer use the unqualified sentence AI has dementia as an academic claim. Dementia is an acquired alteration within an embodied person who has a history, relationships and an established world. Early AI had not lost that world. It lacked persistent worldhood from the beginning.

I would instead describe early generative AI as producing a dementia-like interactional structure or an architecture of cognitive discontinuity. The comparison applies locally to particular characteristics:

• fragile continuity between interactions
• rapid context loss
• locally coherent but globally unintegrated statements
• confabulatory completion when information is missing
• repetition as a means of preserving salience
• dependence upon another participant to restore context and coherence

This reframing also clarifies my role. I was not simply prompting or operating a tool. I was holding the conversational world together. In psychoanalytic terms, there is an echo of holding, although this too must be used carefully. The system did not need emotional containment. What I held was the continuity of the interaction: its history, its priorities and the meaning of its terms.

A brief note on collage

Shevlin makes one further comparison that belongs here, although it should not take over this note. He likens earlier language models to the Surrealist exquisite corpse, in which participants contribute fragments while seeing only part of the developing whole. An intelligible composition can emerge without a single participant controlling its total meaning.

This helps describe the composite and discontinuous character of early generative output. It does not prove mentality, consciousness or metarepresentation. Shevlin’s point is that contemporary agents, with greater behavioural stability, access to current information and interaction with other systems, are becoming harder to explain as nothing more than a collage of inherited human patterns. The distinction between collage-like synthesis and integrated mentality will be developed separately in the next Fairy Note.

Phantom Mirror and the return to the human

In Phantom Mirror, the AI does not need to possess metarepresentational consciousness for the method to function. I begin within human language-based metarepresentation. I speak about a moment of danger, reconsider a memory, describe a relationship and formulate language around something not yet fully understood. The model then produces a visual hypothesis from learned visual and linguistic relations.

I encounter that image perceptually and affectively. I refuse it or recognise something within it. Recognition may arrive before I can explain it. I then return to language and make the image and my response available for renewed reflection.

Metarepresentational reflection to visual synthesis to perceptual affective encounter to renewed metarepresentational reflection

The metarepresentational movement belongs securely to the human side of the encounter. Whether any corresponding mentality exists within the machine remains open. My method does not resolve that question. It creates a situation in which different forms of organisation, human reflection, machine synthesis and embodied recognition, become visible in relation to one another.

Conclusion

My early statement that AI had dementia was naive if taken literally, but it contained a serious observation. It identified the peculiar labour required to work with an intelligence that appeared capable but discontinuous. I repeatedly supplied context, restored identity and reconstructed the local world in which the exchange could continue.

Marslen-Wilson’s distinction between immediate perceptual experience and human language-based metarepresentational consciousness now allows that experience to be described more carefully. Shevlin’s graded account of artificial mentality allows me to compare cognitive profiles without claiming that human and artificial minds are equivalent. His use of Capgras shows that even human belief is neither indivisible nor uniformly integrated.

The question is therefore no longer whether AI has dementia. The question is what kinds of continuity, integration and metarepresentational organisation are necessary before we can reasonably speak of a mind, and what happens to the human participant who must continually provide what the system lacks.

The machine did not have dementia. But the relationship taught me that a world does not remain coherent by itself. Someone has to remember it, repair it and hold it together.

Development of the research

This note revises an early anthropomorphic intuition into a comparative account of cognitive discontinuity and relational world-maintenance. It connects lived dementia care with Marslen-Wilson’s developing distinction between perceptual and metarepresentational conscious experience, and with Shevlin’s multidimensional account of belief and artificial mentality. It also clarifies that the generated visual hypothesis enters metarepresentational reflection through the artist’s use of it, not because the model is assumed to possess metarepresentational consciousness.

References

Marslen-Wilson, William (2026). Informal recorded conversations with the author, 26 August, 1 September and 18 September 2026. Developing account not yet reviewed or approved for publication.

Shevlin, Henry (2026). Three Frameworks for AI Mentality. Frontiers in Psychology, 17, 1715835. https://doi.org/10.3389/fpsyg.2026.1715835

Shevlin, Henry (forthcoming). Aeroplanes Also Fly: Analytic Functionalism and the Possibility of Machine Consciousness. Behavioral and Brain Sciences. https://philarchive.org/rec/SHEAAF-7