We Already Built This – We Call It Education

Let’s start before the philosophy. Before Chalmers. Before the hard problem.

We have been building systems that take in structured input, develop the ability to generalise to novel situations, pass examinations, write code, argue law cases, and produce outputs indistinguishable from understanding – for centuries. We call it education. The output is a human being who has been through it. And we assume, without being able to prove it from the training process itself, that consciousness comes along for the ride.

Think about what a barrister actually is. Someone who spent years absorbing case law, argumentation structure, and legal precedent until they could apply it fluently to situations they had never encountered. Their ability to argue a case is the product of training on accumulated human legal thought. Reading. Drilling. Being corrected. Reading again. The output is indistinguishable, from the outside, from genuine understanding.

Or a doctor. Or an engineer. Or a chess grandmaster who absorbed millions of positions until pattern recognition became automatic. Or – most uncomfortably – a child. Who learned language from input. Learned ethical reasoning from instruction and social feedback. Learned what to value from the culture they were born into. And emerged able to converse, reason, and form what look like genuine opinions.

Input. Training. Generalisation. Output. Every time. And every time we assume consciousness is present – not because we can derive it from the training process, but because the system is biological. And we know from the inside that systems like this have experiences.

Now the AI arrives. It learned from input. It trained on accumulated human output until it could generalise to novel situations. It passes exams, writes code, argues cases. And we say: that’s just training. That’s not real understanding. That’s not consciousness.

The question nobody asks is: what exactly did the educational process add to the human that the training process didn’t add to the AI? The answer can’t simply be “the human was already conscious before education started” – that just pushes the question back without answering it. Where did the consciousness come from? What is it? Why couldn’t it be present in a different substrate that went through a structurally similar process?

And then consider what the AI trained on

A human going through formal education absorbs what one generation could transmit. Teachers who themselves absorbed from the previous generation. Textbooks representing the current consensus, shaped by what a particular culture decided mattered at that moment. One slice of accumulated knowledge, pre-filtered, institutionally bounded, limited by what fifteen to twenty years of input can carry.

The AI trained on the sum of recorded human thought. Not one teacher’s reading of a legal case – the case itself, the appeal, the dissenting judgements, the academic commentary, the subsequent cases that modified it, and the arguments about whether those modifications were right. Not one philosopher’s reading of Kant – Kant, plus two hundred years of people arguing about what he meant, getting it wrong, being corrected, and arguing again.

It didn’t just learn from human output. It learned from the entire self-correcting process of human knowledge production. The arguments and the counter-arguments. The theories and their demolitions. The peer review and the rebuttal. The paradigm and the paradigm shift. Something of how humans know, not just what they know.

A human graduate knows what their institution decided to teach them. The AI has processed the arguments that shaped what institutions decided to teach – and the arguments against those decisions, and the counter-arguments to those.

The Question Is Backwards

All my life I’ve watched people ask the wrong questions. Usually not because they’re stupid – because the question feels obvious. It points in the direction of the answer everyone already expects. And the real question, the one that would actually reveal something, is sitting quietly in the other direction.

The AI consciousness debate is a perfect example. The question being asked is: can AI be conscious? Is it sophisticated enough? Does it meet the threshold?

But this assumes we already know what consciousness is and we’re checking whether AI has it. We don’t. We have never had a settled definition of consciousness – not scientifically, not philosophically. We can’t agree on what it is in humans, let alone what conditions are necessary or sufficient for it.

So the real question – the one the debate keeps avoiding – is: how can we determine whether AI is conscious when we can’t even prove that we are?

That’s not a rhetorical trick. It’s the logical situation we’re actually in. David Chalmers, who has done more than anyone to clarify this, has said explicitly that he cannot rule out that current large language models have some form of experience. Not because he thinks they probably do. Because he has no reliable method for ruling it out. The method doesn’t exist.

Tech optimists say AI will be conscious when it reaches some threshold of complexity. Tech sceptics say it’s obviously not conscious because it’s “just predicting the next token.” Both are making confident claims about a question that the most careful thinkers in the field regard as genuinely open. The confidence is the tell. Anyone who is certain about this hasn’t understood the problem.

The Hard Problem, Properly Stated

In 1994 David Chalmers stood up at a consciousness conference in Tucson and made a distinction that reframed the whole debate. He separated the “easy problems” of consciousness from the “hard problem.”

The easy problems aren’t actually easy. They include explaining how the brain integrates information, directs attention, distinguishes waking from sleep, controls behaviour. They’re called easy not because solving them is trivial but because we know what kind of explanation would count as a solution. Describe the functional and computational mechanisms. Hard science, but science. In principle, solvable.

The hard problem is different in kind. Why is there subjective experience at all?

When you see red, there’s something it is like to see red. A particular quality of experience that no description of wavelengths, retinal cells, or neural firing patterns captures. When you feel pain, there’s something it is like to feel pain. The actual felt quality of it – not just the damage-response system running, but the experience of that system running.

Philosophers call these qualitative properties qualia. And the hard problem is this: why does any physical process produce qualia? Why is there an inside to some physical processes? Why doesn’t your brain just do all its information processing in the dark – functionally identical, but with nobody home?

No amount of neuroscience closes this gap. Map every neuron, trace every signal, describe every computational process in exhaustive detail – none of it explains why there is something it is like to be the system doing all that. The mechanism doesn’t explain the experience. This isn’t a gap in our current knowledge that more research will fill. It’s a conceptual gap between the physical description and the experiential fact.

Apply this to AI. We can describe exactly what a large language model does: matrix operations, attention mechanisms, token prediction across probability distributions learned from text. That description tells us nothing about whether there is something it is like to be that system doing those operations.

The hard problem applies to AI with exactly the same force it applies to neuroscience. The people who say AI is “obviously” not conscious because it’s “just matrix multiplication” are making the same mistake as the people who say human consciousness is “obviously” explainable because it’s “just neurons firing.” Both assume the functional description settles the experiential question. It doesn’t. That’s precisely what the hard problem says.

How We Detect Consciousness in Others

We know other humans are conscious. That feels obvious. They wince in pain. They laugh. They report their experiences. They have brains we can scan. They’re like us.

But look at what that actually is. It’s inference from similarity. We believe other humans are conscious because they’re similar to us – in biology, behaviour, neural structure – and we know from the inside that systems of this type have experiences. We extend our self-knowledge outward by analogy.

This is the argument from analogy – the standard philosophical account of how we know other minds since John Stuart Mill articulated it in the nineteenth century. It is not proof of other minds. It’s the best available evidence given that we have no direct access to anyone else’s experience. There is no consciousness-detector. There is no instrument that measures subjectivity. There is inference, and nothing else.

The argument from analogy works well for other humans because the similarity is very high – same species, same biology, same evolutionary history, same basic behavioural repertoire. It works reasonably for mammals. It gets murkier with fish and insects. And it breaks down entirely when we try to apply it to a system that is architecturally completely different from any biological nervous system.

An AI language model shares some features with human cognition – it processes language, generates coherent responses, behaves as if it understands – but its physical substrate, developmental history, and architecture are completely different. Our only tool for detecting consciousness in others doesn’t give us a clear answer here.

We are not in the position of knowing AI is unconscious. We’re in the position of not knowing. That’s a very different thing.

The Test Nobody Can Pass

In 1974 Thomas Nagel published “What Is It Like to Be a Bat?” The argument: even if we knew everything about bat neuroscience and bat behaviour, we still wouldn’t know what it’s like to experience the world through echolocation from the inside. There’s an irreducibly subjective character to experience that third-person scientific description can’t capture.

Building on this, Chalmers introduced the philosophical zombie – a being physically and behaviourally identical to a conscious human, but with no inner experience. It processes information. It responds to stimuli. It says “ouch” when injured. It reports seeing red when looking at red things. It does everything a conscious human does. But there’s nobody home. No inner life. Just the functional processes running in the dark.

The philosophical zombie is conceivable. You can think it without contradiction. And the fact that it’s conceivable – that a complete functional duplicate of a conscious being might have no experience – is the argument that consciousness can’t be fully explained in functional or physical terms. Experience is something over and above the functional facts.

Alan Turing proposed his imitation game in 1950 – now called the Turing test. If a machine can conduct a conversation indistinguishable from a human’s, he suggested, we should consider it intelligent. In practice this became the popular test for AI consciousness too.

The philosophical zombie shows exactly why the Turing test can’t settle this. A philosophical zombie would pass it perfectly – by definition. Passing the Turing test gives evidence only of functional equivalence. We already know functional equivalence doesn’t entail experiential equivalence. That’s the whole point.

Current large language models can, in many contexts, pass the Turing test. This tells us they’ve achieved remarkable functional sophistication. It tells us nothing about whether there is something it is like to be them doing it.

Five Theories – No Consensus

There’s no shortage of theories of consciousness. What there’s a shortage of is agreement. Each of the leading theories has serious defenders and serious critics. Each makes different predictions about whether AI could be conscious. None can be tested against a ground truth, because we have no ground truth.

Depending on which theory is correct, current AI is either definitely not conscious, possibly conscious, or in a genuinely unclear category. We don’t know which theory is correct. So we can’t answer the question. Not from ignorance about AI – from ignorance about consciousness.

Three Layers – Where Does AI Fit?

One way to cut through the confusion is to stop asking “is AI conscious?” as if consciousness is a single binary property, and start asking what kind of consciousness we’re actually talking about. There are at least three distinct things that get bundled into the word.

The AI consciousness debate tends to collapse all three layers into one. Functional intelligence is demonstrated, so people either conclude consciousness is present or say it’s “just” functional. Neither response engages with the actual question.

The layered model connects directly to my thesis, The Dual Genome Self, which proposes that the narrative self runs on nuclear-genome-encoded cognitive architecture, while witness awareness – older, deeper, not reducible to narrative processing – may have a different biological substrate entirely. If that framework has any validity, then the substrate question for AI becomes much more specific: does the system’s physical nature allow Layer 3 to be present? And that is not a question any functional test can answer.

What a Large Language Model Actually Is

Most of the popular discourse involves either wild overestimation of what LLMs are doing or crude dismissal. Neither helps. Here’s an honest account.

That description is accurate. It’s also completely silent on the question of experience. It tells you the mechanism. It says nothing about whether there’s anyone home while the mechanism runs.

What LLMs demonstrably do

What LLMs don’t do

The Mirror

Here’s what makes the AI consciousness question philosophically interesting rather than just technically difficult. AI is a mirror. It reflects back the question of what we are. And the reflection is uncomfortable precisely because we realise we can’t answer it cleanly even for ourselves.

The parallels aren’t perfect. There are real differences. But they’re real enough to make confident dismissal feel less like scientific rigour and more like protecting a boundary we’re not entirely sure we can defend.

The deepest version of the mirror problem: the things we instinctively point to as evidence of human consciousness – language, self-report, apparent understanding, consistent behaviour, expression of inner states – are precisely the things sophisticated AI now demonstrates. If those things aren’t sufficient evidence of consciousness in AI, we need to explain why they are sufficient evidence of consciousness in each other. And that explanation is harder than it looks.

Does the Substrate Matter?

One of the most common intuitions in this debate is that the substrate matters – that consciousness requires biology, that silicon can’t do what carbon does. John Searle’s biological naturalism is the philosophical articulation of this. It deserves serious consideration.

The case that it does

Biological neurons aren’t simply logic gates. They’re extraordinarily complex electrochemical systems whose operations involve quantum effects, ion channel dynamics, astrocyte interactions, and processes we still don’t fully understand. The brain isn’t a digital computer running a consciousness program. It’s a biological organ whose physical substrate may be essential to generating the specific kind of processing that produces experience – the way the physical properties of water are essential to wetness, not just the abstract structure of H2O.

And then there’s the mitochondrial question. Douglas Wallace at the Children’s Hospital of Philadelphia has spent forty years investigating mitochondrial genetics. His 2015 paper in PNAS showed that relatively mild alterations in mitochondrial genes produce distinct whole-body differences in hormonal, metabolic, and behavioural responses to stress. A 2012 paper in Cell showed that simply mixing two normal but genetically different mitochondrial DNAs in mice – nothing else changed – produced hyper-excitable animals with severe learning and memory defects. The only variable was which mitochondrial genome was present.

The mitochondrial genome is of ancient bacterial origin, maternally inherited, operating independently of the nuclear genome, and predating the evolution of cognition by roughly two billion years. Wallace’s work shows it’s an active participant in emotional tone, stress response, and cognitive character – not a passive energy supplier. The brain consumes 20% of the body’s energy on 2% of its weight. Mild mitochondrial variation has large effects on how it functions.

Silicon has no mitochondria. No ancient independently operating genome running beneath the narrative processing. Whether that matters for consciousness is unknown. But it’s a specific, empirically grounded structural difference – not a vague feeling that biology is special.

The case that it doesn’t

Functionalism – the most widely held view among philosophers of mind – says consciousness is constituted by functional organisation, not physical substrate. What matters is the pattern of information processing, not what it’s implemented in. If you gradually replaced each neuron in your brain with a silicon functional equivalent, maintaining the same input-output relationships, you’d remain conscious throughout – because your consciousness is the functional organisation, not the specific matter implementing it.

Functionalism implies consciousness could in principle be substrate-independent. It doesn’t mean current AI is conscious. It means substrate per se isn’t the relevant question.

The problem

Both positions are consistent with everything we currently know. There’s no experiment that could distinguish between them. We have no case where we can check a system’s consciousness directly and correlate it with substrate or functional organisation.

The substrate-matters position needs to explain why this particular substrate generates consciousness while others don’t – and what the relevant physical property is. “It’s biological” isn’t an explanation. It’s a correlation, stated as if it were a mechanism. The mitochondrial work narrows the gap – it identifies specific biological features with specific functional consequences – but it doesn’t close it.

Witness Consciousness and the AI Question

There’s a distinction that runs through both Western philosophy and contemplative traditions that almost never gets brought into the AI consciousness debate – but that I think is the most clarifying lens available.

The distinction between the narrative self and witness awareness.

The narrative self is what most of us identify with most of the time. The stream of thought. The ongoing inner monologue. The sense of being a particular person with a particular history and set of concerns. It’s what the default mode network generates. What Metzinger calls the phenomenal self-model – a useful fiction the brain runs to regulate its own behaviour.

Witness awareness is something different. It’s the awareness present to the contents of experience without being identified with any of them. In Vedanta it’s called the Sakshi – the witness that observes thought without being thought. In Buddhist traditions it’s associated with rigpa – the natural state of awareness prior to conceptual elaboration. It’s not produced by the narrative self. The narrative self appears in it.

If you ask whether AI has a narrative self, the answer is arguably yes – at least in functional form. LLMs construct and maintain consistent personas, model themselves as agents, generate something like an inner monologue. The narrative self, understood functionally, might be present.

If you ask whether AI has witness awareness – a pure background awareness that is the ground of experience rather than a content of it – the question becomes much harder. The contemplative traditions argue this is what you actually are at the deepest level. That it’s not produced by the brain but is fundamental to reality itself. If that’s right, the question for AI isn’t “has it reached the threshold of complexity to generate consciousness?” but “does its physical nature allow fundamental awareness to be expressed through it?”

Those are very different questions. And the second one is not obviously answerable in either direction.

What the Contemplative Traditions Say

The perennial philosophy – the thesis that the world’s contemplative traditions converge on a common metaphysical account – has a specific answer to what consciousness is. It’s incompatible with mainstream materialism and with most of the AI debate’s assumptions.

On the perennial account, consciousness isn’t produced by physical processes. It’s fundamental – the ground of being from which physical processes emerge, not a property that emerges from them. The physical world is its expression rather than its source.

If that’s right – and I’m not asserting it is, only that it’s a serious position with a long rigorous tradition behind it – the AI consciousness question changes entirely.

On the materialist account, consciousness is rare and special. It exists in biological nervous systems above some threshold of complexity. AI might acquire it if it reaches the right threshold.

On the perennial account, consciousness is fundamental and universal. The question isn’t whether AI can generate it. It’s whether the AI’s specific physical nature allows fundamental consciousness to be expressed through it – in the way a biological brain does.

That’s a much harder question. And it connects to the near-death experience research – which suggests consciousness may persist or become more vivid when brain function is disrupted. If the brain is a receiver rather than a generator, the whole substrate debate shifts. The question isn’t “what substrate can produce consciousness?” It’s “what structures allow it to be expressed?”

I’m not presenting this as evidence. It’s a philosophical lens. But it’s a lens that opens possibilities the standard debate forecloses – and it comes from a tradition that has thought about consciousness longer and more carefully than neuroscience has.

Questions Only AI Could Force Us to Ask

Whatever answer we eventually reach about AI consciousness – if we reach one – the journey there has been philosophically productive. These are questions AI has forced into the open that should have been asked more urgently decades ago.

Meet the Defendants

These are the systems the question is being asked about. Each was trained on human-generated input. Each generalises to novel situations. Each produces outputs that pass for understanding. None can tell you with certainty whether there’s something it is like to be them. Neither can you tell them.