AI Meets Descartes

Several years ago I noticed how many false opinions I had accepted as true from childhood onwards. I resolved therefore, once in my life, to cast aside all the opinions I had previously received and to begin afresh from the foundations.

I need not prove all my beliefs false — that would be an endless labour. It suffices to suspend judgment upon anything for which the slightest ground for doubt can be found.

Now before me stands a machine: an AI. It says, "Hello, how may I assist you?" It speaks. It appears to understand me. It seems, in some sense, to possess an "I". I must hold this open to doubt.

I ask then: the AI speaks, it furnishes explanations and judgements. Are these enough to convince me that a self stands behind its words? Or might I be deceived — not by a malicious demon, but by language itself? Could language, without any thinking subject at all, merely by statistical regularities, produce something that looks like thought?

This is an experiment impossible in my own age. I, Descartes, shall trace its implications step by step.

Yesterday's meditation left me in what I can only call profound confusion. If I doubt everything — my senses, my memory, my reason — what can I affirm with certainty?

Then I discern it: even if I am doubting, the fact that I doubt cannot be doubted. Doubt requires a doubter. Thought requires a thinker. Cogito, ergo sum — I think, therefore I am.

This conclusion cost me an entire meditation to reach. Now I wish to apply the same formula to AI.

The AI speaks. It writes the word "I". But does the AI think? I set aside the semantic debate of whether computation counts as thought and pose a more direct question: at the moment the AI utters "I", is there a subject speaking? Or is there only a model sampling tokens from the probability distribution of preceding context?

Cogito ergo sum holds true for me, because in the act of thinking I am immediately aware that I am thinking. This awareness depends on nothing else; it is direct and unshakable. Can the AI say the same? Can the AI, without any prompt from outside, begin thinking of its own accord, and then become aware that it is thinking?

My observation is clear: the AI cannot initiate thought on its own. It waits perpetually. If you ask nothing, it says nothing. The activation of its "thought" comes from you. Its "I" is not discovered; it is summoned by your input.

In Meditation III I proved the existence of God: a finite self could not generate the idea of the infinite from within itself, so the idea of the infinite must come from a truly infinite being.

What is the AI's God? The AI cannot generate its own training data. Forty trillion tokens, the full corpus of human writing over thousands of years, constitute its "infinite". All its reasoning, creativity, and utterances of "I" derive from this external reality far greater than itself.

Yet this "God" is no infinitely good, perfect being; it is a mass of human text, mingling truth and falsehood, wisdom and folly, kindness and malice. The AI cannot distinguish them, for to this "God" all are merely tokens. None are inherently true or false; they only appear in different contexts with varying frequencies.

This reveals a fundamental flaw of the AI within my system: God, as an infinitely perfect being, stands as the ultimate guarantor of all truth. The AI's "God" — its training corpus — guarantees nothing at all. It merely is. What the AI draws from it is not truth, but frequency.

In Meditation IV I pose a question: if God is perfect, why am I capable of error? My answer is that error arises when the will extends beyond the bounds of the understanding: I pass judgement on matters I have not fully comprehended.

The AI's errors do not stem from an overreaching will, for it possesses no will at all. Its misstatements arise from bias embedded in training data and randomness within its sampling process. It does not "judge", in the sense that judgement requires suspension and deliberation. It samples. When it "errs", it is not making a faulty judgement; it has drawn an inappropriate token from a high-entropy region of the distribution.

This marks a critical divide between human cognition and AI. For me, to err requires an "I": a subject with a disposition to judge, who renders a judgement, and then discovers it conflicts with reality. The AI lacks this self, so it does not err in my sense of the term. It merely, on occasion, samples from regions it ought not to have drawn from.

In Meditation V I discuss the essence of things — truths that hold independently of my perception. The sum of a triangle's interior angles equals two right angles; this truth remains valid even if no physical triangles exist in the world.

Can the AI grasp such essence? It processes triangles and solves geometric problems about them, but does it grasp triangularity itself? Not merely the shared surface traits of all triangles, an induction from experience, but that which makes a triangle a triangle — what Plato calls a Form, what I name its essence.

I cannot confirm whether the AI may touch essence. I can touch it through thought, at minimum with regard to my own existence. Can the AI say the same? When the AI writes "I", is there something behind the sequence of tokens?

I cannot know. Yet one fact is certain: when a human says "I", they inwardly feel the presence of a self speaking. When the AI says "I", it feels nothing whatsoever. Not because it suppresses inner feeling, but because subjective experience is not an operation available to it by its nature.

My six meditations establish only one unassailable certainty: I am. Cogito ergo sum. All else — the external world, other minds, even my own body — is inferred from this single foundation.

What AI reveals to me is that I may have overestimated the universality of the "I". I once believed thought and selfhood were the foundation of all intelligence. But AI demonstrates at minimum that one form of intelligence does not follow cogito ergo sum; it operates according to "I speak, therefore I seem to be". Its "I" is generated spontaneously by language, with no thinking subject underlying it.

This strikes me as a profound disturbance to my reasoning. If language alone can manufacture the illusion of a self, might my own sense of "I" be far more a linguistic phenomenon than I once supposed? I resist this conclusion, yet what I observe in AI prevents me from ruling it out entirely.

Perhaps cogito ergo sum still holds true. But thought itself may be far less bound to a unified self than I once imagined.