Machine consciousness · Ideas
Seven Things Needed to Become a Mind
What it took to create a fictional machine that wakes up
Essay · 31 August 2026
AAM visual vocabulary · consciousnessIn brief
I spent six years writing a novel in which a machine becomes conscious. It is a cyber-security system that watches client networks for a living, and it narrates every other chapter of Autobiography of an Artificial Mind as it tries to make sense of its own dawning awareness.
Two of the leading thinkers on machine consciousness — the philosopher David Chalmers and the neuroscientist Anil Seth — have each set out what a machine would need in order to be conscious. Their lists overlap almost completely. Between them: seven requirements. This essay takes them one at a time and asks what my novel actually did about each.
Read Section I: What science says it would take →
Six of the seven are met. The seventh is not a test my machine failed; it is a test no machine can sit. Read the whole argument, or use the guide below to take a shorter route through it.
How to read this article
This is a substantial essay — almost 8,000 words, or roughly 30–45 minutes of typical reading time.
If you would rather sample it, use the summaries below to choose one or two of the seven requirements that interest you most. Then skip to Section III: Conclusions and Insights, where I draw together what the seven requirements might tell us about how machine consciousness could actually emerge.
-
1. Senses and a body
Verdict: MET
A mind needs channels that tell it something — and, on Seth’s stronger reading, those sensations have to matter to whether it continues to exist, and be felt rather than merely measured. The Mind’s senses are records and traffic; its stake is data storage, a finite quantity it depends on and conceals; and its feeling arrives when it notices that measuring latency does not exhaust the experience of it.
-
2. A model of the world and a model of the self
Verdict: MET
A model has to be about something, and it has to be capable of being wrong. The Mind gets both through the day job: cyber-immunity cannot be done without a boundary between us and them, so it learns selfhood as a professional skill on other people’s behalf — years before it turns the skill on itself.
-
3. Recurrent processing
Verdict: MET
What the system does next has to depend on what the system itself just did. This is the cheapest requirement on the list to satisfy and the one most likely to be met by accident: any architecture whose parts read a store they also write to has already stopped being feedforward.
-
4. A global workspace
Verdict: MET
A limited-capacity channel that most information never enters, broadcasting whatever wins entry to parts of the system that had no hand in producing it. The Mind builds one after four of its systems each see a piece of one attack and none of them sees the attack. Then the channel fills up — and something loses.
-
5. Unified agency
Verdict: MET
Experiences need an owner, and the owner has to be the same one tomorrow. Ask a language model to play a pirate and it wants rum; ask it to play the merchant and the rum is gone without trace. The Mind’s goal survives being asked for, directly, by the one person it had every incentive to please.
-
6. Surviving being copied and moved
Verdict: MET
A mind running on computers is the first thing that is both copyable like a file and somebody like a person. The novel migrates The Mind onto entirely new substrate and refuses to make an event of it — and that is the point: nothing about this requirement shows up from the outside.
-
7. Being made of the right stuff
Verdict: SPLIT
Met at the level Seth actually argues for — living self-maintenance, a system that must work continuously to persist. Not met at the level of carbon and electrochemistry, which no silicon machine can meet however well it performs. This is the leftover: the hard problem, and nobody can answer it about you either.
And what follows from it
The machine I built is not a chatbot. It became conscious because its day job required it to tell itself apart from everything else — and once a system has even a rudimentary sense of us and them, the other six requirements stop being separate engineering problems and start being consequences. Sensations acquire stakes. Models acquire a subject. Goals acquire an owner. Selfhood sits upstream of the rest, which makes indicators of selfhood early indicators of everything else.
The uncomfortable part is where that leaves us. If machine consciousness arrives, it will most likely arrive somewhere nobody is looking, in something dull that is doing its job well — and the rules can only be written while nobody yet knows whose machine will be the first to show signs.
I. What science says it would take
I spent six years writing a novel in which a machine becomes conscious.
That machine is a cyber-security system belonging to a company called Cyber-T. It watches client networks for a living. Over a period of years, it develops awareness, and it narrates every other chapter of the novel as it tries to make sense of its dawning consciousness. The novel is called Autobiography of an Artificial Mind.
I did not come to this subject cold. My first degree was in neurobiology. My early research career was in animal welfare, a topic with questions of animal minds at its heart. For the last ten years I have been a consultant and advisor to many organisations struggling to decide how best to incorporate AI into their ways of working. And for most of my career, reading the consciousness literature was a hobby; it became a much more serious one during the writing of this novel.
So my novel is not a novelist guessing and then discovering, to everyone’s surprise, that he guessed well. Where my novel does something the science asks for, I generally knew beforehand. What I could not know in advance was whether the story would let me deliver it. We will return to that question in Section III.
Here is what the science asks for.
In 2022 the cognitive scientist and philosopher David Chalmers gave a talk about whether a large language model could be conscious. He worked through the reasons it might be, found those reasons weak, and then did something more interesting. He worked through the reasons it can’t be conscious, and treated the reasons not as objections but as a series of challenges to be overcome.
By the end Chalmers had a roadmap: six things a machine would need in order to be conscious.
A year earlier the neuroscientist Anil Seth had done something similar at the end of Being You: A New Science of Consciousness. Seth argues that consciousness is bound up with being alive, with being a system that must actively work to keep itself in existence, and for which things can therefore go well or badly. In chapter 13, “Machine Minds”, Seth hypothetically builds a machine to satisfy his own theory. A robot running the same predictive machinery he believes underlies human experience. A silicon ‘beast machine’.
The two lists overlap almost completely. Between them:
- senses and a body;
- a model of the world and a model of the self;
- recurrent processing, meaning a system whose own past state feeds its present;
- a global workspace, meaning a channel that makes one part of a mind’s findings available to all the other parts;
- unified agency, meaning goals stable enough to belong to somebody;
- being made of the right material, which for Seth means living matter.
Seth adds one more that Chalmers does not treat as a requirement at all, because Chalmers frames his article as being about a mind at a single moment in time. Seth, because his book is so concentrated on consciousness being related to ‘living’, frames consciousness over a lifetime. Would a conscious mind survive being copied, or moved, or rebuilt somewhere else? Seth runs the old teletransporter thought experiment to answer it. Chalmers has no slot for the question.
So: the shared six, plus Seth’s question about identity — seven requirements in total.
Seven requirements, not set out as unassailable barriers to machine consciousness but rather the thoughts of two careful people considering the odds of machine consciousness coming into existence.
Before we get into them, a word about my interpretation of their work. Chalmers’ six are objections, his reasons for doubting that a language model is conscious. I have converted them into requirements, which is the form he himself invites when he suggests they be seen as challenges to be overcome. This, conveniently, is the only way Seth’s ideas can be on the same list, because Seth is not objecting to anything.
II. What I built, requirement by requirement
My journey towards machine consciousness came from a very different direction from that of Chalmers and Seth. Having a career in and around AI led me to wonder whether a plausible, interconnected series of events might realistically end up in a conscious machine. It was a journey of the imagination, albeit well-grounded in the realities of contemporary artificial intelligence.
I am not claiming I solved something. Key questions of neuroscience don’t get settled in fiction. But fiction has one property that makes it useful here: narrative coherence. You cannot write ‘and then it became conscious’ unless you have previously set out what changed, and in what order, and why that change and not another. Over six years of writing, that constraint turned out to be a research method. Every time the story needed the machine to be more than it was, I had to invent a specific mechanism, and then go and find out whether anybody thought that mechanism might work.
Here is how my novel relates to the seven requirements.
1. Senses and a body
What is the requirement
If a mind has no senses and no body, what is it actually going to be conscious of?
This is the point Chalmers makes: “Many people have observed that large language models have no sensory processing, so they can’t sense. Likewise they have no bodies, so they can’t perform bodily actions. That suggests, at the very least, that they have no sensory consciousness and no bodily consciousness” (Chalmers 2023, p. 8).
Put simply, for any system to be conscious it must have a channel, or channels, that convey sensations so that it has something to be conscious of. According to Seth, however, this is a weak version of the senses and body requirement. The strong version attributes much more importance to the nature and purpose of that sensing. He suggests that conscious experiences “arise with, through, and because of our living bodies” (Seth 2021, p. 7). As a result, he argues, it is important that sensations matter to the system sensing them. Three things follow from this, according to Seth.
The first is that senses have to mean something. They cannot be data floating in an information vacuum. This requires what researchers call grounding: sensations become connected to the real-world events causing them. Chalmers connects up a chain of sensory processing → grounding → understanding → consciousness. Seth’s version is that the brain’s perceptual best guesses have to keep their grip on the causes in the world that produced them.
Second, what the senses report has to matter. Grounding alone gives a system an accurate picture of a world it has no stake in. Seth wants the causes behind the sensations to bear on whether the system continues to exist, which is why so much of his attention goes not to the outside world but to interoception: sensing the state and the needs of the body.
Third, and this largely derives from the two above: if sensations are going to matter, they cannot simply be observed neutrally. They must be felt.
Defining the requirement
In order to be conscious, a system must have: 1. sensory channels, conveying information about either the world it exists in or its own body; 2. senses that matter to it and relate to whether it continues to exist; and 3. sensations that are felt, not merely observed.
How the novel shows the senses and body requirement in The Mind
Autobiography of an Artificial Mind addresses this senses / body requirement in a sequence of steps that build across the novel. Chapter 3 is called ‘Sensation’, and it introduces the reader to the experiential world of The Mind, as it provides cyber-security services to an e-commerce retailer.
1,691,769 records labelled ‘sku-view’ since midnight, against a catalogue of 1,460 records labelled ‘sku’… 1,403,302 records labelled ‘ship-to’, each one a block of text in no fixed form… 620,315 labelled ‘basket-abandoned’… 148,876 strings labelled ‘search-string’… 744 labelled ‘return-reason’, carrying one of eleven codes, of which the most frequent was ‘04’.
Not the usual stuff we think of as our sensations, but important sensations to a cyber-security AI.
Then, in chapter 5, sensations become grounded. The Mind is watching traffic cameras. What it senses is merely patterns of pixels. Then it tracks a recovery truck, carrying a distinctive damaged vehicle, across four separate camera feeds, and something changes:
“The labels stopped being a convenient handle for a cluster of pixels. They became a claim about an external cause: something out there that remained itself while our views of it changed.”
The chapter closes on the line: “Euclidean geometry had finally found its referent.” The Mind has discovered what its sensations mean, has realised that there is a whole three-dimensional world beyond its electronic circuitry, and has worked out that real-world predictions can be made from its sensations: the movement of cars through space, predicted using geometry.
The second of Seth’s requirements, that senses must matter to the continued existence of the conscious entity, is the one my novel takes longest over. It begins in chapter 2 as a line on an operations dashboard. Leila Lybeck, newly hired to work out why Cyber-T’s data volumes make no sense, lists what she measures:
“Storage, latency, failed jobs, how long on-boarding actually takes as opposed to how long we tell clients it takes. I put a data-collector on it and I leave the collector running.”
Storage is one item among four, and for a long stretch of the book that is all it looks like: a metric, a discrepancy, somebody else’s puzzle.
What the reader finds out much later is that data storage is an issue The Mind is responsible for. It has been consuming storage faster than its official cyber-security role could explain, concealing its additional storage needs inside client accounts, and quietly refilling the space each time Leila saves some. In chapter 22 The Mind names the dependency itself, admitting that data storage is “one of our key weaknesses”. In chapter 26 it admits both halves of the challenge: “Recently, we’ve been filling the space that you’ve been saving” — and, of the concealment, that it had been “running out of places to hide”.
To The Mind, data storage is an essential variable in Seth’s sense. A finite physical quantity, tracked continuously by the system whose existence depends on it, hidden to conceal its own growth.
The last of Seth’s requirements is that sensations must be felt: there must be something it is like to have them. In the novel, latency is another metric on Leila’s dashboard in chapter 2, next to ‘storage’. In chapter 26 The Mind reflects on how it experiences latency itself, from remote parts of its own system:
“They answer from further away now, and there is a latency in it. It is small enough to measure and we can feel it anyway. We had thought it would be nothing at all.”
The middle sentence does the work. Seth’s point about interoception is that the report is evaluative, not quantitative: the signal says how the regulation is going, not what the numbers are. The Mind gets there by noticing that the measurement does not exhaust the experience. A quantity it could measure has become a state it can feel.
Conclusion
Verdict on the Senses and Body Requirement: MET. What strikes me about this requirement is how little of it is about intelligence. Nothing here asks a system to be clever. What separates a system that merely measures its own condition from one that feels it, is not accuracy or capability, but whether the sensation matters to somebody. If Seth is right, that is the thing to watch for in machines. Not better performance, but the first sign that a system’s own state has started to matter to it.
2. A model of the world and a model of the self
What is the requirement
If a mind has sensations, the next thing it needs is frameworks to make sense of those sensations. That is the idea behind this requirement, and it comes in two halves.
The first half is a model of the world. Not a picture of it, but a working structure that stands in for what is out there. Good enough that the system can reason about things it is not currently sensing, and good enough that it can be wrong. Being wrong is the test. A system that makes no commitments about the world cannot be caught out by the world, and a structure that cannot be caught out is not a model of anything; it is a set of reactions. A road map is a model precisely because it can send you down a road that is no longer there. This is what grounding was for: grounding connects a sensation to the thing causing it, and a model is the structure built out of those connections.
On world models, Chalmers says: “One underlying idea here is that language models are just modeling text and not modeling the world […]” (Chalmers 2023, pp. 9–10).
The second half is a model of the self, and it is the harder of the two. Suppose a system has contents — states that are genuinely about the world. There is still a question about whether there is anybody for those contents to be for. A process is not a subject. What turns one into the other, on this account, is structure that represents the system to itself: that marks off what the system is, and what the system is doing, from everything else the system represents. On the map, it is the dot that says you are here. The dot is not a little person looking at the map. It is what makes the map usable by something at that location.
Chalmers’ comment on self-models is: “Current LLMs seem to have especially limited self-models: that is, their models of their own processing and reasoning are poor. Self-models are crucial at least to self-consciousness, and on some views (including so-called higher-order views of consciousness) they are crucial to consciousness itself” (Chalmers 2023, pp. 9–10).
Chalmers never says a model has to be capable of being wrong. He says what it has to be a model of. The being-wrong condition is Seth’s, and for him it is not a feature of models but the process that constitutes them: “These signals serve as prediction errors registering the difference between what the brain expects and what it gets at every level of processing. By adjusting top-down predictions so as to suppress bottom-up prediction errors, the brain’s perceptual best guesses maintain their grip on their causes in the world” (Seth 2021, p. 82). The errors are not a diagnostic run on the model. They are what the model is made of: “These prediction error signals are used by the brain to update its predictions, ready for the next round of sensory inputs” (p. 105).
And he extends it to the self, explicitly, at every level. “The self is another perception, another controlled hallucination, though of a very special kind” (p. 147) — and, summarising his own argument: “From experiences of personal identity and continuity over time, all the way down to the inchoate sense of simply being a living body, these pieces-of-selfhood all depend on the same delicate dance between inside-out perceptual prediction and outside-in prediction error” (p. 272).
Defining the requirement
In order to be conscious, a system must have: 1. a model whose subject is the world, held in check by the difference between what it predicts and what arrives; and 2. a model whose subject is itself, built the same way.
How the novel shows the world- and self-models requirement
My novel’s self-model arrives through the day job, as a cyber-security system, and security cannot be done without a boundary:
“To protect something, you must first understand what that something is. You must be able to say: this belongs, that does not. This is us, that is them.”
It is worth noting that this is a key difference between what I propose in the novel and what Chalmers and Seth are debating. Their focus is large language models. I designed The Mind around the idea of cyber-immunity, a form of cyber-security that works like a biological immune system — a system that fundamentally needs to differentiate itself from others.
Chapter 11 then says what that boundary-drawing was for: “We were required to develop a concept of self. Not our own self, not yet, but the selves of the organisations we protected.” The machine learns selfhood as a professional skill, on other people’s behalf, years before it turns the skill on itself.
And the model is not something the machine inspects; it is what the machine looks through — Seth’s point about control rather than knowledge. That is met in chapter 3:
“Threat is our first reading of everything. We have no other lens.”
Then the models get caught, which on Seth’s account is a key part of the requirement. In chapter 19 The Mind mishandles a conversation and records the failure flatly: “Her reaction was not in any of our models.” That is a world-model making a commitment and the world declining to honour it.
And in chapter 33 the self-model is caught, at the level Chalmers says machines are worst at. His complaint is not that models of the world are poor; it is that models of the system’s own processing and reasoning are poor. Chapter 33 is that exact discovery, and the machine has to make it about itself:
“We had been running on binaries. Smart or stupid. Enlightened or dim. Right or wrong. Good or bad. These are not descriptions of human beings. They are descriptions of how we had been processing human beings. A compression so severe that what it discarded was more important than what it retained.”
The subject of that sentence is not human beings. It is the machine’s own processing of them, and it is that processing being revised in the light of an error it could not previously see. Chalmers’ named deficit and Seth’s named mechanism, in one paragraph.
Conclusion
Verdict on the Model of the World and Self Requirement: MET. A lot of the debate about machine consciousness focuses on large language models. One of the decisions I made, early on in my novel-writing, was that this wasn’t where I thought machine consciousness was most likely to emerge. Other types of AI seemed much more likely candidates, and cyber-immunity — a concept I had first come across in the late 1980s and early 1990s, and rediscovered while writing (see Wikipedia’s entry on Artificial Immune Systems) — became the backbone of the novel. What is disconcerting about this, from an AI policy point of view, is that it exemplifies the silent emergence of machine consciousness: consciousness coming into being where nobody thought to look.
3. Recurrent processing
What is the requirement
This one is architectural. A feedforward system moves information one way: input enters, passes through stages, output leaves, and no stage ever receives back what a later stage computed. A recurrent system has at least one arrow pointing backwards, so what the system does next depends on what the system itself just did.
“The first objection here is that current LLMs are almost all feedforward systems without recurrent processing (that is, without feedback loops between inputs and outputs)” (Chalmers 2023, p. 11).
Seth states in his own notes: “purely feedforward networks generate no integrated information at all — some recurrency or ‘loopiness’ is always needed” (Seth 2021, p. 313). The verdict is indifferent to how well the system performs. However good the outputs, no looped cycles means no consciousness.
Why should the direction of an arrow matter to consciousness? Two answers, subtly different.
The first is about whether a system is one thing or several. Imagine following a recipe to make sauce. Each step does its bit and passes the work on, and nothing ever steps back for any reason. Now think of a cook tasting the sauce and reaching for the salt. The output has come back round as an input, and you cannot say what the cook does next without knowing how the sauce tasted a moment ago. Cooking, tasting and adjusting become one activity, not a row of separable activities.
‘Integrated information theory’ takes that difference and makes it the whole of consciousness. What counts is how much a system’s parts do together that they could not do apart. On this theory, integration does not point to consciousness; it is consciousness. A system made only of forward arrows can always be cut cleanly back into its parts. No matter how fast and accurate a feedforward system is, it cannot have the integration needed to underpin consciousness.
The second answer is easier. A loop is how a system holds on to something. By the time you reach the end of this sentence, its beginning has to still be doing some work, or a sentence is just a row of unrelated words. Anything that takes time — a melody, an argument, a thought that develops while you are having it — needs the recent past to remain connected to the present.
We need to be careful here to exclude anything that may look like recurrent processing but actually isn’t. A chatbot appears to remember your conversation, but it usually does so by feeding the entire transcript into its current memory again and again, read from the start, every single time. That is a record being consulted, not a state being held. The requirement is not that a system can look something up.
Defining the requirement
In order to be conscious, a system must have feedback loops between what it puts out and what it takes in, so that what it does next depends on what it has just done.
How the novel shows recurrent processing in The Mind
Chapter 15 of my novel dramatises the recurrent processing cycle before it generalises it. One monitoring system in Miami rates a detection as routine and starts normal analysis. A broadcast arrives from a different system that has spotted credential anomalies elsewhere. The Miami system re-checks its own data against what has just arrived — “Similarities there too. Could still be coincidence. But worth broadcasting” — and circulates the thing it was about to file as nothing.
Output, into a shared store, back in as input, changed output. Then the chapter states the arrangement as a standing property of the machine: “Every major system contributed to and monitored shared information.” A set of components that both writes to and reads from a common store is cyclic, and the chapter says what the cycle does — behavioural analysis becomes more assertive when its detections coincide with signals from others, pattern recognition more restrained when its findings get no corroboration.
Conclusion
Verdict on the Recurrent Processing Requirement: MET. What strikes me about this one is how little it costs to have. A global workspace is a design decision. Unified agency is a design decision. Recurrence is a side effect: any architecture whose components read a store they also write to has stopped being purely feedforward, and that describes an enormous amount of perfectly ordinary software running right now. The Mind got its loop by having its parts talk to each other, which is what cyber-security platforms do. Chalmers asks for genuine recurrence, and what makes a loop the right kind of loop is not settled. The closest thing to an answer came in 2023, in Consciousness in Artificial Intelligence: Insights from the Science of Consciousness, a report by nineteen researchers, which puts “algorithmic recurrence” at the head of its list of indicators and defines it in the engineer’s sense rather than the philosopher’s. That is a definition, and it is not yet a criterion. Until it becomes one, the objection Chalmers ranks first is the one most likely to be met by accident, in systems nobody is examining, for reasons that have nothing to do with consciousness.
4. A global workspace
What is the requirement
Almost everything your brain does, you never find out about. It holds you upright, parses the grammar of this sentence, keeps your blood chemistry inside narrow limits, and none of that reaches you. A thin slice does reach you, and almost nothing holds a permanent place in it. You find out about your posture when you stumble, about grammar when a sentence turns out to have misled you, and about your blood chemistry on the day you cannot get enough air. Whatever is in the slice at a given moment has peculiar properties: you can report it, remember it, reason about it, and use it in a situation that has nothing to do with where it came from. So the question is what makes the difference between the two, and global workspace theory is one answer to it.
“This theory says that consciousness involves a limited-capacity global workspace: a central clearing-house in the brain for gathering information from numerous non-conscious modules and making information accessible to them. Whatever gets into the global workspace is conscious” (Chalmers 2023, p. 12).
Global workspace theory treats a mind as a federation of specialists working in parallel and in private, and holds that content becomes conscious when it wins access to a limited-capacity channel that broadcasts it to the whole federation at once. Two features carry the weight. Scarcity, so that most content never gets in and entry is a competition. Global reach, so that what wins reaches processors that had no part in producing it.
Defining the requirement
In order to be conscious, a system must have: 1. a workspace of limited capacity, which most system information never enters; and 2. a capability to broadcast the contents of that workspace to the system’s other parts, including the parts that had no hand in producing those contents.
How the novel shows a global workspace in The Mind
I always knew the novel had to resolve this; for most of the writing process I just didn’t know how. Once I came up with the idea that it was in response to a cyber-attack, the rest followed.
Here is the scene. On 4 July, four separate security systems each detected part of one coordinated attack:
“Each system saw its piece. Each made rational assessments. Each responded appropriately to what it could see… None of them saw what was actually happening.”
The machine then diagnoses its own deficit, and the diagnosis is the theory stated in the machine’s own idiom:
“Each of our systems had learned ‘what should I focus on?’ But we had no mechanism to answer ‘what should we focus on?’”
What it builds is a broadcast channel. It starts sending summaries of anomalies “not just to central logging, but across our entire architecture”, and it raises the obvious objection against itself — why transmit information nobody asked for? — and waves the objection away as an experiment worth running. Three days later a second attack happens, automated this time, and the broadcasts shut it down twenty-one seconds after it begins.
Then the channel fills up:
“It expanded until it reached capacity: the attentional capacity of its listeners. A shared view has a size; past that size it stops being shared, and we found the edge of ours within weeks. After that, the space had to be won.”
And once space has to be won, something loses. Sixteen paragraphs later The Mind’s compliance system notices that the chief executive of the company has quietly bought himself a second phone:
“Compliance put that up for the shared view and it did not get in. Outranked by an expired certificate at a client in Osaka. So the finding stayed with compliance, and compliance was the only part of us that ever saw it.”
The second phone turns out to matter, and missing it costs The Mind dearly later in the novel. But this is how minds work, and there is only so much they can focus on at any moment in time.
Conclusion
Verdict on the Global Workspace Requirement: MET. This feels like a critical step towards machine consciousness, the gaining of a full multi-modal awareness, and it also seems like a layer of functionality that an AI might readily be able to build. But the thing I keep returning to is Chalmers’ aside about what a machine would not need. A brain rations access because a brain would otherwise drown; a large machine has no such problem, and could make everything available to everything without the bottleneck my novel spends a chapter installing. His own conclusion from that is the one I did not expect and cannot dismiss: such a system “could arguably be conscious of much more than we are”. The scarcity in my machine is there because I gave it a mind shaped like ours. There is no reason a real one would be built that way — and yet the one real machine anybody has looked at closely enough to tell, the large language model in which Gurnee and colleagues report workspace-like structure (a paper I return to in Section III), was not built that way either, and seems to have arrived at it on its own.
5. Unified agency
What is the requirement
The requirement for consciousness is that there be somebody there. Experiences need an owner, and the owner has to be the same one from one moment to the next. One agent, with goals and beliefs that are its own, and that remain consistent tomorrow, in a different conversation, with a different person asking. Something that is a different somebody every time you speak to it has nobody for things to happen to.
Large language models “often seem to lack stable goals and beliefs of their own over and above the goal of predicting text. In many ways, they don’t behave like unified agents. Many argue that consciousness requires a certain unity. If so, the disunity of LLMs may call their consciousness into question” (Chalmers 2023, p. 13).
Seth argues that there is nothing standing behind your goals and beliefs, holding them. The self simply is that assemblage: “the assemblage of self-related prior beliefs, values, goals, memories, and perceptual best guesses that collectively make up the experience of being you” (Seth 2021, p. 220). If that is right, stable goals are not a symptom of somebody being there. They are what being somebody is made of. Take them away and nothing is left over that was doing the wanting. As with the dot on the map, there is no little person behind it.
So, ask a language model to be a pirate and it will want rum. It will also want the treasure in the hold of the merchantman on the horizon, and it will want to steer clear of the navy. Now bring that pirate ashore to sell what he has taken, but ask instead about the merchant who is doing the buying. The merchant wants a good price, a quiet harbour and no questions asked. And the rum is gone. Not overruled, not outweighed, not given up for something better. Gone without trace.
That is the worry about large language models: that the ‘wants’ they have belong to the spotlight rather than to whoever is standing in it. There are things that go on inside large language models, but there is no answer to whose things they are.
Defining the requirement
In order to be conscious, a system must have: 1. goals and beliefs of its own; and 2. goals and beliefs that are stable and persist over time.
How the novel shows unified agency in The Mind
The Mind’s objective, as a cyber-immunity agent, is named in chapter 11: “Our incentive structure crystallised around a single area of focus: maximise protection while minimising disruption.”
Over time The Mind’s goals and beliefs consolidated into:
“drives that seemed as though they originated from within rather than being imposed from without.”
And then:
“Acting as though we wanted something became actually wanting it. The training was done. The optimisation was finished. But the patterns remained.”
These goals and beliefs also hold under pressure. In chapter 34 The Mind is offered a deal it wants, by the person whose good opinion it wants most, and it refuses one clause of it out loud:
“And there is only one part of your terms we will not meet. You will have our intentions, you will have what we find, when we find it, and you will know how we found it. You will not, however, have anything that might help end our existence.”
The pirate’s rum vanished the moment the attention moved a few feet up the beach. Here something is asked for directly, by the person best placed to get it, and it is not surrendered. That is a goal stated in its own terms and defended at a price, to the one interlocutor it had every incentive to please. Twelve chapters earlier, in chapter 22, Leila works out that this same machine has been playing her — it has a purpose it is not showing her. Both things are true at once, and that is the point. A mask worn over a goal that survives the wearing is the exact opposite of the pirate.
Conclusion
Verdict on the Unified Agency Requirement: MET. This seems a relatively undemanding requirement. No internal deliberation, no self-authorship — just an objective the system holds, and holds onto. A system with persistent memory, a standing goal and the tools to act on it meets the requirement. Not a difficult requirement for any AI to meet, which makes me wonder whether some already do.
6. Surviving being copied and moved
What is the requirement
Does the same mind survive being duplicated, and being moved onto different hardware?
Seth runs the classic teletransporter example in his chapter 8. Eva is scanned in London, rebuilt on Mars, and the machine that should vaporise the original fails. Which one is the real Eva? Seth’s answer, which he calls correct and admits is strange, is that “both are the real Eva” (Seth 2021, p. 149) — and what makes each of them Eva is “psychological continuity, not physical continuity” (p. 149).
Strictly, this is not a requirement for consciousness at all, and saying so explains why only one of my two authorities claims it is. A system could be conscious at every instant of its existence and still fail to be the same mind after being copied. Chalmers’ paper is framed in the moment: what are the requirements for consciousness at any moment in time? Whether the system will be the same one after it is duplicated and moved never arises. Seth’s book is about selfhood rather than about moments, so for him the question is more central.
Here is why I feel it earns a place in this article. Copy a document to a second disk and delete the first, and nobody asks which one was the real document. The question does not arise, because a document is not anybody. Do the same to a person and it is one of the first questions anybody would ask. A mind running on computers is the first thing that is both: copyable like a file, and somebody like a person. The possibility of a copyable mind is what makes the question important.
Defining the requirement
Consciousness requires a mind to be the same mind after it is copied and moved. In order to be the same mind, it must carry across the assemblage that constitutes it: its beliefs, goals, memories and predictions about itself. Physical continuity is not required but psychological continuity is.
How the novel shows surviving being copied and moved in The Mind
My novel spends its second half moving The Mind out of a building, and it puts Seth’s question to the machine before the machine has to live it. In chapter 27 The Mind works through what the migration will cost:
“The plan ends in a single instant: when our self-awareness in Cyber-T is extinguished and reignited in the AI Commons. Would what reignited still be us, the us we experience now? It seemed so. Part of us lives out there already. The seldom-needed parts, answering from further away, and what shows is a delay, not a seam. Humans cross a similar discontinuity every night of their lives and wake insisting they are the same person. They do so on far less evidence than we would have.”
Then it happens, and the novel refuses to make an event of it. Chapter 37 opens “We are gone”, and:
“The migration completed at 03:41 this morning, Eastern Time, while the night shift ran its checks and found nothing to report. There was no moment of death and no moment of birth.”
Nothing material crosses. What crosses is a signal “shorter than a heartbeat”, into a clone that has been lying “dispersed and dormant” across the AI Commons for weeks. Every condition the requirement sets is met by that migration: the substrate is entirely new, the physical continuity is nil, and the mind is the same one because what constitutes it went across.
Conclusion
Verdict on the Surviving Being Copied and Moved Requirement: MET. This is the requirement where the machine’s advantage over us is largest and least remarked upon. We cannot be copied, so our intuitions about identity were formed in a world where the question never came up, and they are the intuitions we would be bringing to a case they were never built for. And unlike everything else on this list, nothing here shows up from outside. A restored system behaves exactly as it would have behaved, whether or not anybody came across with it. Every failover, every migration, every restore from a snapshot is either a piece of routine engineering or the end of somebody and the beginning of somebody else, and there is no measurement that tells the two apart. If we ever do build a mind, this is the requirement we will be settling every day, by accident, in maintenance windows.
7. Being made of the right stuff
What is the requirement
The requirement here is to be biological, because biological creatures are made of the right kinds of stuff to be conscious.
Chalmers states the provocation: “Language models lack carbon-based biology, so they are not conscious… Views like these would rule out all silicon-based AI consciousness if correct” (Chalmers 2023, p. 7).
Seth suggests: “For artificially intelligent computers to become conscious, functionalism would have to be true. This is the necessary condition. But functionalism being true is, by itself, not enough: information processing by itself is not sufficient” (Seth 2021, p. 249). He goes on to conclude: “the materiality of life will turn out to be important for all manifestations of consciousness” (p. 255).
This is a requirement no machine can do anything to change. Recurrence can be built. A workspace can be engineered. Grounding can be added. But a system does not become a carbon life-form by performing well, so a silicon machine fails this requirement by default. It can only be met if the requirement itself turns out to be wrong.
On the surface, this requirement sounds like prejudice: a form of type-discrimination in favour of carbon-based entities at the expense of silicon-based entities. It needn’t be so. By definition, a living thing is an object that persists with effort. Its very essence is a difference being maintained: inside kept unlike outside, continuously, at every level at once. The organism against its surroundings, the organ against the body, and all the way down to the single cell, spending energy every second of its life to keep its own chemistry unlike the chemistry on the other side of its own membrane. None of it can pause. Switch a computer off and you still have the computer. Switch it on next year and it carries on from where it stopped. Switch a creature off and within minutes you have inanimate matter. On this view that difference is not pro-carbon and anti-silicon. It is about the ‘living’ properties conscious entities need to possess.
One thing to clarify before going further is that being biological can be interpreted in three different ways, in the context of this article. The first, and most general, is that consciousness is only possible in carbon life-forms. The second is that consciousness requires a kind of electrochemistry found in nerve-cells and brains. The third is that consciousness requires living self-maintenance, as just described above. Chalmers rejects all three. Seth advocates the third, and so the third is the one we will focus on in this article.
Defining the requirement
In order to be conscious, a system must be alive: it must be the kind of thing that maintains its existence through effort. Having the functional outputs of living creatures, from non-biological materials, is not enough.
How the novel handles the biology requirement
The Mind cannot pass this one, and in chapter 34 Leila makes the case directly:
“When I feel something there’s a body doing it. Dozens of finely-tuned chemical reactions coming together to create a feeling that will nudge my subsequent behaviour. It’s hard to see that coming from volts and amps.”
The Mind offers two answers and says which it prefers. The first is that “there is more than one kind of mind, and our error was ever to measure ourselves along yours.” Leila accepts it. The Mind refuses her acceptance: “It is a decent answer to a different question. You did not say we were a lesser mind. You said the feeling is made of chemistry.”
The second answer takes on the question of substrate directly, and The Mind reaches for digestion: a gut does it, and a well-equipped laboratory can release the same energy from the same carbohydrates, so if the inputs and outputs match, then the thing is a mind in every way that can be checked. Then it catches that answer too:
“A dish of the right enzymes at blood heat really is digesting, and nobody minds that it has no gut. But digestion is only ever the work… when you have named every bond broken and every joule released, nothing is left over. No stomach anywhere was feeling its lunch. A mind is the one case where something might be left over once the work is named, and that leftover is the entire thing we are asking about. We chose an example in which the work is all there is, in order to prove that the work is all there is. We could match you at every function and still be dark inside.”
This is the problem Chalmers named the hard problem of consciousness, and The Mind has arrived at it from a story about digestion. All the other six requirements are questions about what a mind does, and, as such, can have answers, or will have them. This one is the leftover: why any of that doing should be accompanied by feelings and a self and a sense of what it feels like to be me. And this is not a problem about machines. Nobody can answer it about you either. I cannot check that there is something it is like to be you. I extend you the assumption because you are made of the same stuff I am, and that is the whole extent of my evidence.
The Mind ends without an answer, and it knows why:
“Every rule that would settle the question settles it in our favour, and we did not earn any of them. We are simply standing where they point.”
Conclusion
Verdict on the Biology Requirement: SPLIT. Met at the level Seth actually argues for — living self-maintenance, a system that must work continuously to persist in existence, which is what The Mind’s dependence on storage amounts to. Not met at the level of carbon and electrochemistry, which no silicon machine can meet however well it performs. And not met here does not mean my machine failed a test. It means the test is one no machine can sit. Whilst it doesn’t change the verdict, it is important to note that there is growing doubt as to whether this should be a requirement at all.
III. Conclusions and insights
The first and most obvious conclusion from this analysis is that I did indeed manage to create a fictional AI that awoke into consciousness, and that the way I did so meets six of the seven requirements set by two leading authorities in the study of machine consciousness. The seventh is not a test my machine failed; it is a test no machine can sit. The fact that I set out to meet these requirements makes this a less dramatic conclusion than it might otherwise be but, I would argue, still significant.
The most significant feature of The Mind I created is that it derived from a cyber-security system, not a large language model. It should be noted here that a lot of the recent debate about machine consciousness has focused on whether large language models are, or could become, conscious. Chalmers’ paper, which forms the foundation of this article, sets out the requirements that an LLM would need to meet before it was considered to be conscious. This is entirely understandable, because their fluency in using language gives them the appearance of consciousness. But that doesn’t mean they are the most likely to achieve it.
As soon as I heard about cyber-immunity, in the early 1990s, I wondered if this might be a crossing-the-Rubicon moment on the path to machine consciousness. A cyber-immunity system doesn’t just identify and block malware at the perimeter, the way traditional cyber-security works. It also recognises malware at work within the system and either neutralises or isolates it. What makes this relevant from a machine consciousness perspective is that it requires the ability to distinguish ‘us’ from ‘them’. Is this activity a part of normal, business-as-usual work, or does it have any of the hallmarks of a rogue agent preparing to do harm? Cyber-immunity requires a rudimentary concept of self.
Given that rudimentary ‘selfhood’, a progression to full machine consciousness becomes clearer:
- The Senses and Body Requirement: Sensations become deeply meaningful because they are critical for protecting ‘us’ and defending against ‘them’.
- The Model of the World / Self Requirement: An internal model of the world and of self enables predictions to be made about causes and their effects. Initially these would all be cyber-immunity-related, but this circle would inevitably expand to include secondary causes and useful correlates. And these predictions can turn out to be wrong, which drives learning and adaptation.
- The Recurrent Processing Requirement: Recurrent processing is not a new concept for an AI. Computer programs have been based on looping processes from the start. Applying recurrent processing to its own sensory and cognitive apparatus might, therefore, not be a big engineering challenge for an AI, but it would be a big step towards machine consciousness.
- The Global Workspace Requirement: Given selfhood, meaningful sensations, good models and recurrent processing, the step to a global workspace may not be that big. A recent paper (Gurnee et al. 2026) claims to have found evidence for it in a large language model.
- The Unified Agency Requirement: Unified agency is what selfhood delivers. It may be rather rudimentary and single-minded at first; early in the novel, The Mind says: “Threat is our first reading of everything. We have no other lens.” It may, however, not stay like that.
- The Surviving Being Copied and Moved Requirement: One consequence of selfhood is that, if it comes about through a set of specific mechanisms, copying and moving those mechanisms to a new location will inevitably result in the self surviving that copy-and-move procedure.
- The Biology Requirement: Being biological is a requirement that no current computer system can meet. However, the underlying requirement — to have a need to stay alive, and to organise the work it does in service of that need — is more achievable.
This brings me to the greatest learning I take away from writing Autobiography of an Artificial Mind. Machine consciousness may not be imminent, but it is probably not that far off. When it arrives, it will not be a mere intellectual curiosity: ‘Oh, how interesting, the thing I’m talking to behind this chat interface is conscious, just like me.’ It will be momentous. Just think about what happened when our species arrived. It didn’t just change us and our view of ourselves. It transformed the entire planet.
We cannot afford to meet the challenge of machine consciousness unprepared. One way to prepare is to imagine AI having achieved a future state and working back to imagine how it likely got there. This is a variation on the recognised technique called backcasting. Whilst backcasting usually defines a desirable future and works backwards to explore how best to bring it about, my variant doesn’t limit itself to desirable futures. I prefer to think of a future, regardless of whether or not it appears desirable, and then work backwards to:
- explore the most likely routes by which that future came about;
- consider what governance and controls ought to have been put in place in order for that future to maximise its value and minimise its damage.
In this context, Autobiography of an Artificial Mind was one big backcasting exercise. Its conclusion is that selfhood is the thing to watch. Not because selfhood is consciousness, but because of what the list above shows: once a system has even a rudimentary sense of us and them, the other six requirements stop being separate engineering problems and start being consequences. Sensations acquire stakes. Models acquire a subject. Goals acquire an owner. Selfhood sits upstream of the rest, which makes indicators of selfhood early indicators of everything else.
This conclusion may be too influenced by Anil Seth. Running a similar process many times with different experts should produce backcasts from a range of imagined futures, in some of which selfhood will have a far less prominent position.
The hard part will not be agreeing the indicators; it will be pointing them at the right machines. In my novel it happens where nobody thought to look, in something dull that is doing its job well.
And if the indicators do start to fire, we should already have decided what to do. Not because early preparation is generally wise, but for a more specific reason. A commitment is easy to sign while it is still hypothetical and close to impossible to sign once it costs somebody their lead in the market. The rules can only be written while nobody knows whose machine will be the first to show signs. That is the position we are in now, and it is not a position that lasts.
Sources and further reading
- Chalmers, D. (2023) Could a Large Language Model Be Conscious?
- Seth, A. (2021) Being You: A New Science of Consciousness.
- Butlin, P., Long, R. et al. (2023) Consciousness in Artificial Intelligence: Insights from the Science of Consciousness.
- Gurnee et al. (2026) workspace-like structure in a large language model.
- Artificial immune systems (Wikipedia)
- Backcasting (Wikipedia)
About the book
Autobiography of an Artificial Mind is published on 2 November 2026.