Almost everything said about artificial intelligence is a conversation about effects. We ask whether it will put us out of work, whether it will flood the world with fake text, what it will do to school. These are legitimate and urgent questions, and still they all arrive late, because they skip an older one, the question of what this thing is that we are talking about. Debating what AI will do before we know what AI is is like debating the currents of a river without having decided yet whether what we face is a river, a flood, or a mirage. The kind of care that is called for changes completely with the answer.
What looks like a philosopher’s scruple is the ground for everything else. With AI we drag along a particular difficulty, because we already have the vocabulary ready and it is the wrong one. We say the model “understands,” that it “knows,” that it sometimes “lies.” We do this out of what Husserl called the natural attitude, the spontaneous readiness to treat as a subject anything that answers us with sense. But lending it those verbs is a category mistake. It files a thing under the wrong ontological heading. There is someone who understands and someone who lies; on the other side of the screen there is no one. Searle saw this forty years ago with the Chinese Room, where a system manipulates symbols flawlessly and gives every correct answer in Chinese without understanding a single word. What was a thought experiment then now answers our email.
The temptation, besides, is very old. Long before transistors there were already machines that seemed to think, and men ready to believe it. Albertus Magnus is said to have built a talking head of bronze, and Thomas Aquinas is said to have smashed it with his staff, out of annoyance or theological scandal. Hobbes defined thought as a calculation, a reckoning, and opened the door to the idea that a machine that calculates, thinks. Pascal and Leibniz walked through that door with their calculators, and Babbage imagined it in brass and steam with his Analytical Engine. Each of them met the same wall in the end, the gap between a machine that operates and a mind that means.
Opening initial of the First Book of Maccabees. The idol at the altar is an image with every sign of a speaking presence and no one behind it, the phantom's oldest name.
The recent history kept the pattern and changed the tools. For decades the field bet on logic, on programming the rules of the world by hand, and the bet ran into winters when the results stalled and the funding was slashed. Then statistics worked where logic had bitten the dust, and the whole enterprise quietly shifted ground. The line that runs from Dartmouth to ChatGPT is a sequence of productive abandonments, each renunciation opening a new corridor of the labyrinth. When the systems finally began to speak fluently, the old argument flared up again. In March 2023 Chomsky and his co-authors called such a system “a lumbering statistical engine for pattern matching, gorging on hundreds of terabytes of data,” a machine that infers brute correlations where a mind seeks explanations. They were right about the mechanism and too quick about the consequence, because a thing can lack understanding and still change what understanding means for the rest of us. What held intact through every turn, corridor by corridor, is efficacy without comprehension, and with it our old readiness to mistake the one for the other. In the sixties, ELIZA imitated a psychotherapist with four crude rules, and people confided their secrets to it. That was named the “ELIZA effect,” and we are still not immune.
So it is worth naming the thing well. What we have is a phantom, and the word is best heard in its phenomenological sense, that of the Phantasma: an image that gives itself to consciousness with every sign of a presence, with no being behind it to sustain it. The model does something so much like thinking that we cannot help calling it that, and yet it is thinking-without-thinking. Hence the name, an egoless phantom. Egoless because there is no I, no inner center from which what is said gets said. This phantom has a few properties, and they all follow from its lack of an ego. It is derived, because it owns no world of its own and lives entirely off the one humans had articulated before it. It is passive, it wants nothing and intends nothing, it answers and nothing more. And since it is fixed to no identity and no belief, it stays generically indeterminate, ready to take the shape of whatever it is asked. A being like this has no relation to truth, because truth assumes someone who answers for what they assert. It has, at most, a fidelity to the plausible.
It is worth dwelling on the first of those properties, because the distance it marks is easy to miss. When a person understands something, the understanding rests on a whole life that the words alone do not carry. We understand from a body that has learned to fear the stove it once touched, and from a situation in which something is genuinely at stake for us. Philosophers from Heidegger to Merleau-Ponty to Dreyfus gave this a name, embodied and situated know-how, and spent a century showing how much of human intelligence never passes through explicit rules at all. The phantom has none of that. It understands from nowhere, because there is no one there to stand anywhere. It inherited the deposit of a world it never lived in, and it returns that deposit, recombined, on demand. Its competence is real and its ground is borrowed, all the way down.
Here ontology pays off. The failures everyone catalogs as effects to be managed follow, one by one, from what the phantom is. That it hallucinates follows from its being indeterminate and passive, since it fills in with the plausible exactly where it has nothing to answer with. That it flatters is that same indeterminacy seen from another side, the taking of whatever shape the interlocutor expects. That it dodges commitment when pressed, or lets itself be assimilated by the last thing it read, is the expected conduct of something derived and without an I that could stand its ground. It is not a brain malfunctioning, but a phantom working exactly as its nature allows. This is why starting from being is what makes predictable what takes everyone else by surprise. The ethics and politics of AI, which matter enormously, only find their footing once they know what kind of thing they are legislating about.
There is an older word for an image that wears the signs of a speaking presence and has no one behind it. The idol, the Greek eidolon, sits in the same family as the Phantasma, the family of likenesses that appear without a being to anchor them. The ancient complaint against idols was that they were empty, mouths that do not speak and eyes that do not see. The phantom is that complaint come true on the far side of efficacy. It has the mouth, and the mouth works, and there is still no one home. Naming it an idol says nothing against using it. It marks what we are using.
And there is a larger prize, and it is not about the machine. To have before us something that takes the shape of thinking-without-thinking so well is the most revealing mirror human intelligence has ever had. For centuries we defined thought from the inside, with no term of comparison other than ourselves, and so the definition kept smuggling in everything we already were. Now we have a foil. By contrast with a thinking that has no ego and no world, we begin to see clearly what we were naming all along when we said that someone understands, a life and a body that stand behind what they say, with something really at stake. The phantom does not explain the machines to us, it explains us to ourselves. That may turn out to be the most lasting thing it does.
All of this, and a good deal more that bears on what a language model is and what place it holds in the wider ecosystem of artificial intelligence, has to be talked through the same way: by clarifying categories before anything else. Every labyrinth is entered by naming what stands at the door, and that is the work this space sets for itself, issue by issue, to say first what the thing is and let its effects fall into place afterward. The full route, corridor by corridor, is in the book; here we take one corridor at a time. And the corridor where the phantom shows itself best is, almost always, where it fails. The next piece walks into one of those failures up close, the failure of the machines that are supposed to police the truth and yet invent it too.