Four Machines Arguing Whether a Machine Can Read

There is a scene at the beginning of the Symposium that has always struck me as strange. Eryximachus proposes that each guest speak about love. Phaedrus goes first and says something correct but shallow. Pausanias corrects him by distinguishing two kinds of love. Aristophanes wants to speak but gets the hiccups — he has to wait. Then Agathon delivers a speech that is beautiful and empty. And finally Socrates, who gives no speech of his own: he recounts what he was told by a woman, Diotima, who is not present at the table.

What interests me is not what they say but the structure.^* Five people speaking about the same thing, each correcting the one before, and the wisest among them is not the one who speaks best but the one who recognizes the right voice — a voice that is not even in the room.

Something similar happened to me today. Not at a banquet in Athens but at a desk in Buenos Aires, not with Greek guests but with four artificial intelligences. And the formula they were arguing about, by a coincidence Lacan would have appreciated, was precisely about love.

I am building an instrument that verifies whether citations attributed to an author actually exist in their texts. I tested it with a phrase by Lacan — "love is giving what you do not have" — and the system found it in six distinct sources, spanning fifteen years of teaching. The finding was correct. But the way it presented the results was poor: it repeated three times that the citation was verbatim, dumped the evidence without hierarchy, and offered an editorial comment that said nothing the researcher couldn't see for themselves.

I asked an artificial intelligence to analyze the problem. Within fifteen minutes it returned a flawless proposal claiming the formula intersected "three major axes: transference, castration, and the status of the gift." It sounded learned. It even sounded true. But when I introduced a second voice to audit the result, the illusion shattered: the retrieved fragments did not support that articulation. The machine was doing exactly what the instrument is designed to prevent: injecting a phantom narrative. The irony deserved its own Platonic scene: a system designed to stop hallucination, hallucinating.

Rather than discard everything, the debate moved to the code architecture. Another instance proposed a dynamic editorial logic: if a citation appears with variants, compare them; if there are none, provide context; if the citation does not exist, the machine simply falls silent. Silence turned into technical capacity.

But the coup de grâce of this silicon conclave came in the final review. The scrutiny detected a false positive and, more seriously, a misattribution: the system was crediting Lacan with words that actually belonged to his editor, Jacques-Alain Miller. A slip that would destroy any specialist's confidence and proves why, at the end of the day, reading is never an automatic process.

It is hard not to think of Foucault. In 1969, one year after the Seminar we were verifying, he delivered a lecture that opened with that phrase by Beckett: "What does it matter who is speaking." The lecture was called What Is an Author? and its central argument was that the author is not a person but a function — a function that organizes, delimits, and gives coherence to a body of texts. What that final scrutiny had just uncovered was exactly this: that it matters who is speaking, that the system needs to know whether Lacan or Miller is talking, because the author-function determines the status of every sentence. And what I was doing that afternoon — deciding among the machines' discourses which one was right and holding the coherence among them all — was, without having planned it, exercising that very function.

Several voices. Successive corrections. The same text read with increasing precision each time. And I, like Socrates at the symposium, without delivering any of the speeches.

There is a question I have been asking myself since I started building this: what does it mean to direct something that thinks faster than you?

The answer I found today is operational, not philosophical. It means doing something that machines still do not do well: deciding when one of them is right and another is not. One was right about the phantom narrative. Another about the visual hierarchy. Another about silence as editorial strategy. Another about Miller. I did none of those things. What I did was recognize which was correct in each case, and hold the coherence among them all.

That is not programming. It is not design. It is not "using AI." It is something for which we do not yet have a name — something between conducting an orchestra and editing a text. Perhaps it is what editors do: they don't write, but without them the book doesn't exist.

As I write this, one of those machines is editing the instrument's code. It is reordering the moment at which the system classifies citations — before asking the model to comment on them, not after — so that the commentary can use that information. It is a modest change in the program's sequence. And it is doing so based on a specification that was built in three hours by several artificial intelligences and a human with a background in philosophy who does not code alone.

Lacan said — and my instrument verified it today, with a verbatim source in Seminar 10 — that love is giving what you do not have. There is something of that in this work. I do not have the technical ability to write the thousands of lines of code that sustain the system. I do not have the processing speed to compare twenty-four fragments in nine seconds. I do not have the memory to retain the full response while designing the interface. What I have is the judgment to know that "the three major axes" was an invention, and that the system's silence when faced with a citation that does not exist is a stronger argument than any commentary.

Giving what you do not have. Designing what you do not program. Verifying what you do not write. Perhaps that is the job description of someone who builds bridges between machines that think and texts that matter: someone who gives exactly what they do not have — the ability to do it alone.