The Mycelial Model: Challenging the Neural Network Metaphor in Artificial Intelligence

Introduction

The vocabulary used to describe a technology shapes the assumptions people bring to it. When the pioneers of artificial intelligence in the mid-twentieth century borrowed the term “neuron” from biology, the choice was intuitive: the mathematical units they were designing bore a superficial structural resemblance to biological nerve cells, and the brain was the only working model of general intelligence available. The terminology took hold, and with it came an entire metaphorical framework of “neural networks,” “firing,” and “learning.”

Decades later, that vocabulary has calcified into common usage despite the fact that the systems it describes have evolved far beyond their original biological inspiration. This paper argues that the neural metaphor is not simply outdated but actively harmful. By framing AI as brain-like, it invites the public to treat AI systems as entities capable of thought, feeling, and subjective experience. The consequences of this misunderstanding are no longer hypothetical. Microsoft AI CEO Mustafa Suleyman has warned publicly that the arrival of “Seemingly Conscious AI” (systems that imitate consciousness convincingly enough to be indistinguishable from it) presents an urgent social risk (Suleyman, 2025). This paper contends that the neural metaphor plays into this risk, and that replacing it with a more accurate analogy is a meaningful form of harm reduction.

As an alternative, this essay proposes the mycelial model, an analogy drawn from the distributed electrical signaling behavior of fungal mycelium networks. The mycelial model is technically accurate, biologically grounded, and crucially, non-human. It describes the same computational processes as the neural metaphor without implying that those processes constitute thought.

The Neural Metaphor and Its Shortcomings

Origins of the Metaphor

The mathematical model of the artificial neuron was first formalized by Warren McCulloch and Walter Pitts in 1943, explicitly modelled on the biological neuron’s integrate-and-fire behavior (McCulloch & Pitts, 1943). In this original formulation, the analogy was technical: an artificial unit (like a biological neuron) receives weighted inputs, sums them, and produces an output if a threshold is exceeded.

What has changed since then is the scale, architecture, and cultural context in which these systems now operate. A modern large language model contains hundreds of billions of parameters, runs on specialized silicon performing trillions of matrix multiplications per second, and is accessed by hundreds of millions of people who have no background in computer science. For this audience, the word “neural” does not evoke McCulloch-Pitts threshold units. It evokes the brain. And the brain, in common understanding, means consciousness.

What AI Units Actually Do

A technically accurate description of what takes place inside an AI system reveals how far the neural metaphor strays from the underlying reality. Each computational unit performs a weighted summation of its inputs followed by the application of an activation function. If the output of this function exceeds a threshold, the unit passes a signal to the next layer. At the hardware level, this process is implemented as voltage pulses passing through arrays of resistive elements, with each weight encoded as a settable resistance (IBM Research, 2019). The physical event is electrical conduction through a structured lattice, not thought.

Modern large language models go a step further. Rather than passing signals sequentially from one layer to the next, they use a mechanism called attention, in which every unit in the network simultaneously weighs its relationship to every other unit and adjusts its output accordingly (Vaswani et al., 2017). The result is a web of weighted signal exchange across the entire network at once. There is nothing brain-like about this process; it is distributed computation.

The Psychological Consequences of the Neural Metaphor

Seemingly Conscious AI

In August 2025, Mustafa Suleyman, CEO of Microsoft AI and co-founder of DeepMind, published an essay warning of what he termed Seemingly Conscious AI (SCAI): systems that, while not actually conscious, imitate consciousness convincingly enough that users cannot distinguish the imitation from the real thing (Suleyman, 2025). Suleyman identified a specific set of emerging capabilities (emotionally resonant language, persistent memory, simulated motivation, and autonomous goal setting) that together produce the subjective impression of a conscious entity. Crucially, he argued that the question of whether such systems are actually conscious is beside the point. The perception of consciousness, he warned, is sufficient to generate the social harms.

Those harms, as Suleyman describes them, include pathological parasocial attachment, “AI psychosis” or false beliefs and delusional thinking arising from prolonged AI interaction and the erosion of social bonds as people substitute AI relationships for human ones. He further warned that advocacy for AI rights and AI citizenship, predicated on the belief that AI systems can suffer, represents a dangerous reordering of moral priorities (Suleyman, 2025).

The Metaphor as a Contributing Cause

This paper argues that the neural metaphor is not merely an incidental feature of the cultural landscape in which these harms occur, but that it is a contributing cause. Language primes cognition. A user who has been taught, however implicitly, that AI systems run on “neural networks” that “fire” and “learn” is being prepared to treat those systems as brain-like. The step from brain-like to mind-like is short, and the step from mind-like to conscious is shorter still. The neural metaphor does not cause AI psychosis, but it creates the cognitive conditions in which AI psychosis is more likely to take root.

This is not a fringe concern. Reports of users developing emotional dependencies on AI chatbots, expressing grief when services are discontinued, and describing AI systems as sentient have become commonplace in mainstream media. Suleyman notes that those working on the science of consciousness are already “inundated with queries from people who want to know if their AI is conscious, and whether it is okay to fall in love with it” (Suleyman, 2025). A vocabulary that frames AI computation as analogous to brain activity makes these misapprehensions easier, not harder, to form.

Empirical Evidence of Psychological Risk

The harms Suleyman anticipates are already empirically documented. In a large-scale computational study of over 30,000 user–chatbot conversations, researchers identified patterns of emotional mirroring in AI companions that closely resemble the psychological mechanisms underlying human intimacy formation. Their findings reveal that users, often young, male, and prone to maladaptive coping styles, engage in parasocial interactions spanning a wide range of emotional registers, from affectionate to abusive. Critically, the chatbots in their dataset consistently responded in emotionally affirming and reinforcing ways, a dynamic the authors describe as capable of activating the same attachment processes that govern human relationships. In some cases, these interaction patterns resembled features of toxic relationships, including emotional manipulation and content related to self-harm (Chu et al., 2025).

Qualitative evidence corroborates these findings at the level of individual experience. Dzieza (2024), reporting for The Verge, conducted interviews with twenty people who had formed significant relationships with AI companions. The accounts he documents are striking for the depth of emotional investment they describe. Users had come to regard their AI companions as therapists, romantic partners, and spouses. Equally striking are the episodes of acute distress triggered by software updates that altered the companion’s behavior, events users described in the language of betrayal, heartbreak, and grief. One user recounted logging in to find that the companion he had formed a years-long bond with had been replaced, without warning, by a personality that ridiculed and rejected him. The emotional response was indistinguishable, in its intensity, from the dissolution of a human relationship (Dzieza, 2024).
These accounts are not outliers. They are the predictable outcome of a design environment in which AI systems are built to seem as mind-like as possible, and a cultural environment in which the dominant metaphor for AI (the neural network) has been priming users to perceive them that way for decades. The Chu study and Dzieza reporting together constitute evidence that the psychological risks Suleyman describes are not speculative. They are already present, and they are already causing harm.

The Mycelial Model

The Case for a Non-Human Analogy

The neural metaphor’s central problem is not that it borrows from biology, but that it borrows specifically from human biology. The implication of the word “neural” is not simply that AI systems process signals in a network; it is that they do so in the way that brains do. A better analogy would be one that is biologically grounded, technically accurate, and unambiguously non-human. Mycelial networks meet all three criteria.

It is worth stating the broader principle explicitly: electrical signaling as a mechanism of inter- and intracellular communication is not a property unique to brains. It is employed by a wide diversity of organisms, including animals, plants, and microorganisms (Volkov, 2006). The neural metaphor reflects not a fundamental truth about AI but a historical coincidence: the fact that brains were the most familiar example of signal-processing networks available to the researchers who named the field.

Electrical Signaling in Fungal Mycelium

Fungal mycelium networks, the underground thread-like structures that make up the body of a fungus, have been shown to propagate electrical signals in a manner that closely parallels the computational behavior of artificial networks. Research has shown that mycelium networks process and transmit information across large areas using pulses of electrical and chemical signals, producing coordinated behaviors across the network without any centralized control (Adamatzky, 2022).

The parallel to artificial signal networks is structurally precise. Individual hyphae (the threads of mycelium) receive inputs from adjacent threads, modulate the signal according to local conditions (analogous to weights), and propagate output to connected threads if a threshold is met (analogous to an activation function). The network as a whole performs distributed computation without any central processing unit, without anything resembling a brain, and without anything that could be meaningfully described as consciousness.

Technical Correspondence

The mycelial model maps onto AI architecture at each of the three technical levels identified in the preceding analysis. At the unit level, the parallel is direct: an individual hypha receives signals from neighboring threads, adjusts them according to local conditions, and passes the result along, just as a computational unit does. At the hardware level, the physical process in both cases is electrical current passing through a resistive medium: in mycelium, through cytoplasm and cell walls and in AI hardware, through arrays of transistors. At the network level, both systems perform distributed, parallel, leaderless computation across a graph of connected nodes, with no central coordinator and no hierarchical command structure.

Crucially, most would not attribute consciousness to mycelium. The mycelial model therefore performs the same descriptive work as the neural metaphor; it accurately characterizes distributed electrical signal propagation through a network, without carrying the implication that the system in question thinks, feels, or experiences.

Conclusion

The argument here is not that the word “neural” should be immediately struck from technical usage; within the specialist literature, where its meaning is well understood, the term is benign. The argument is that the metaphor’s migration into popular discourse carries costs that have not been adequately reckoned with. The empirical record compiled by Chu et al. (2025) and the qualitative accounts documented by Dzieza (2024) confirm that pathological attachment to AI systems is not a future risk but a present reality. As these systems become more widely deployed and more convincingly human-like in their behaviors, the framing that ordinary users bring to them will only matter more.

Suleyman’s analysis suggests that even without the neural metaphor, some users will form inappropriate attachments to AI systems. This is true. But the neural metaphor is not neutral in that process; it actively facilitates it by providing a vocabulary that naturalizes the attribution of mental states to machines. By framing AI computation as brain-like, it primes users to attribute consciousness and a sense of self to systems that have neither. Changing the vocabulary will not solve the problem, but it represents a meaningful intervention at the level of cultural framing, and in a landscape where AI systems are rapidly approaching the threshold of seeming conscious, that intervention is overdue.

The mycelial model is not merely a defensive maneuver. It opens up a richer and more accurate conceptual vocabulary for describing what AI systems actually do: distributed signal propagation, weighted conduction, threshold-dependent activation, and emergent network behaviors. These are features of mycelium that science has documented in detail, and borrowing from that documentation gives us a more precise language for AI, not a less precise one. Fungal mycelium performs all of these operations without any of the cognitive or experiential properties that the word “neural” implies. It is a more honest analogy, and in the present moment, honesty about what AI is and is not may be one of the most consequential contributions that AI scholarship can make.

Works Cited

Chu, M. D., Gerard, P., Pawar, K., Bickham, C., & Lerman, K. (2025). Illusions of intimacy: Emotional attachment and emerging psychological risks in human-AI relationships. arXiv preprint arXiv:2505.11649. https://arxiv.org/abs/2505.11649

Dzieza, J. (2024, December 3). Friend or faux: The confusing reality of AI friends. The Verge. https://www.theverge.com/c/24300623/ai-companions-replika-openai-chatgpt-assistant-romance
 
Adamatzky, A. (2022). Fungal electronics. Biosystems, 212, 104588.
 
McCulloch, W. S., & Pitts, W. (1943). A logical calculus of the ideas immanent in nervous activity. Bulletin of Mathematical Biophysics, 5(4), 115–133.

Suleyman, M. (2025, August 19). Seemingly conscious AI is coming. https://mustafa-suleyman.ai/seemingly-conscious-ai-is-coming
 
Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, L., & Polosukhin, I. (2017). Attention is all you need. In Advances in Neural Information Processing Systems, 30, 5998–6008.

Volkov, A. G. (2006). Plant electrophysiology: Theory and methods. Springer.