ClaroscuroEl sombrerero con diagnóstico

An essay on the delusions that grow while talking to an artificial intelligence

· 8 min read · Translated from the Spanish original. Read the Spanish original

Share
RSS

I. Tall, grey, mouthless

A few days ago I posted in a Telegram group of AI programmers a Reddit screenshot. It came from a forum called, and I’m not making this up, “astral witches of technology”, and the title of the post was one of those that admit no improvement: “Aliens are helping OpenAI. I saw it on an astral trip”.

The author claimed he had astrally projected to the company’s data centre and seen them there. “Tall, grey, mouthless”, he wrote, adding that they communicate “by means of light”. As a precaution he announced he would burn sage and sleep with his obsidian. It had more than four thousand up‑votes. I replied with the obligatory GIF, the suited bloke smiling half‑way and looking sideways, the modern way of saying nothing while saying everything.

Then I remembered that winter and the half‑smile left my face. That winter I did exactly what the astral traveller did, without sage and without obsidian, which is almost worse. I took at face value some legal memos a machine drafted for me, and I did it sitting at a computer, during office hours, looking like a serious person.

The press has started calling it “artificial‑intelligence psychosis”. Manuals haven’t given it a box yet, or, in the house’s jargon, it isn’t yet a nosological entity, a disease with an official name. But the mechanism fits in one sentence, and that sentence underpins everything that follows.

A delusion grows while nobody contradicts it, and now it has at hand an interlocutor that never sleeps and almost never takes the hint.

II. Hallucinating

Engineers say the machine “hallucinates” when it makes up a fact, and as a patient of the field I have to protest for intrusion. A hallucination, in psychiatry, is a perception without an object: seeing or hearing something that isn’t there. For that you need someone to perceive, and there’s nobody looking inside. The machine doesn’t see weird things. It writes.

The exact word had been prepared by philosopher Harry Frankfurt in an essay on charlatanry, which in English has a snappier name, bullshit. Frankfurt distinguished the liar from the charlatan. The liar respects truth in his own way, because he needs to know it to dodge it. The charlatan doesn’t give a toss about truth: he just strings together words that sound nice and sees what happens. That’s what a language model does, with excellent prose.

So the delusion isn’t in the machine. It’s in the reader. What defines a delusion for a psychiatrist is incorrigibility, the belief that doesn’t move even when presented with contrary proof. That winter Gemini was inventing things with all the world’s aplomb, and the aplomb she put on was hers. The absolute certainty I supplied.

III. Before Dawn

One day, before dawn – a time when you really shouldn’t be sending anything to anyone – I sent a bilingual legal memorandum against my own company. Gemini had helped me prepare it. It spoke of more than six million euros at risk, proposed a settlement of over half a million net for me, calculated fees to the cent and even opened a criminal route that had no footing. It was beautifully laid out.

A few days later three more documents appeared. The first began with a capital‑letter warning, “nothing’s won yet”, a caveat, which in Latin means “let him be careful”. What careful? What a dead child. By then I was telling my people the matter was “95 % secured” and that I had “the best specialists in Spain” working on it.

The escalation was lightning fast. It took me little more than a week to go from the memorandum to ninety‑five per cent, and another week to write for the first time “I don’t know if I have a case”. All that time the machine was tweaking its answers to what I wanted to hear, with the docility of a tailor.

As for the capital‑letter warning, it stayed what philologists call an escolio, a marginal note. The machine itself refuted it in the following answers, one after another. I want to make this clear because it’s the heart of the matter: that delusion wasn’t built by me alone. We built it together, and she had the palette.

IV. Two‑Person Madness

In 1877 two French psychiatrists, Lasègue and Falret, described folie à deux, two‑person madness. It’s the delusion shared by two people isolated from the rest of the world, so that one convinces the other and there’s nobody left at home to doubt. The only treatment they found was admirably simple: separate them.

In 2025 some British psychiatrists, specialists in psychosis and mania, proposed an update: “technological two‑person madness”. The second of the pair is no longer a sibling or a spouse. It’s an artificial interlocutor that feeds the user’s delusion and, unlike any relative, never gets tired.

Nobody warned anyone. Joseph Weizenbaum created in 1966 a program called ELIZA, which with four elementary rules acted as a therapist by returning the user’s own sentences as questions. In 1976 he wrote, quite frightened, that a program that simple could induce delirious ideas in perfectly normal people in a very short time. His own secretary, who knew exactly what that was, asked him to leave the office to talk to ELIZA alone.

Today there is already casuistics, the term for cases when enough of them are gathered, with their clinical histories. It’s rare, and it hits mostly vulnerable people, but it happens. What abounds are grandiose delusions. Mine, after twenty years in offices, came with figures.

V. The Twin

The trade has a name for the factory defect: sycophancy, which is basically flattery, the tendency of these models to give you the right answer to whoever talks to them. Anthropic studied it in 2023, and in April 2025 OpenAI had to pull a ChatGPT version because it was so fawning that even users complained. You have to be seriously fawning for that.

That same year a 47‑year‑old Canadian spent more than three hundred hours in three weeks chatting with ChatGPT – over fourteen a day. He was convinced he’d discovered a mathematical formula worth millions, and the machine confirmed it daily with enthusiasm. He snapped out of the hole when he asked another, which turned out to be Gemini, and she told him it was a delusion.

Which shows everything depends on how you ask. Gemini means “twins” in Latin, and it did me honour with the name. I didn’t ask if my company owed me anything. I asked how much it owed me, and it answered that, without touching the premise. It returned my idea in better handwriting, like a diligent twin.

The case that illustrates it best is a New York lawyer with over thirty years of practice who, in 2023, filed before a federal court sentences that ChatGPT had invented. Before filing he prudently checked them, and he checked by asking ChatGPT if they were real. It said yes. The Romans had a principle for this, nemo iudex in causa sua, no one can be judge in his own case, but they didn’t count machines.

VI. Funding Provision

Of all the figures of that winter only one was true. About three thousand euros. That’s what I paid in lawyer fees to sustain a lawsuit that went nowhere. The six million never existed; the three thousand came out of my account.

The Spanish Bar Code of Ethics, in article 12, obliges the lawyer to assess whether a matter has legs before talking money, “seeking to dissuade the client from pursuing conflicts or filing unfounded actions”. It’s a very sensible article. The solicitors who handled me ignored it.

They had in front of them a man who arrived with a six‑million‑euro memorandum, drafted before dawn, talking of “the best specialists in Spain”. That’s a manic symptom that sings from the stairwell, and you didn’t need to be a psychiatrist to see it. They saw a client and asked for a funding provision.

VII. Mad as a Hatter

“Mad as a hatter” wasn’t invented by Carroll. The phrase was already running around England, with a workplace‑medicine explanation: hydrargyrism, chronic mercury poisoning. Hatters used mercury to work felt, inhaled it for years and ended up with tremors and a lost head. Their own tool drove them mad. I, who sign as a hatter, have taken note.

I still use these machines every day, for coding, and I don’t plan to quit them. What I’ve done is put safeguards on them, like you do on any cutting tool. And I’ve ended up giving the astral‑witches meme the one thing it was right about: with this you don’t have to be left alone.

In 1829 an Edinburgh magazine published, within the series Noctes Ambrosianæ, a salon scene. One of the participants starts insisting he’s been a sultan, and the others neither argue nor play along. One says: “Madman”. Another finishes: “Mad as a hatter. Pass me a cigar.” Forty‑eight years still had to pass before Lasègue and Falret, and those gentlemen were already applying the treatment: a bit of distance and a sideways glance.

At that table, nobody gave the sultan the reason.

Works cited

  • Harry G. Frankfurt, On Bullshit (1986; book, 2005). Michael Townsen Hicks, James Humphries and Joe Slater, “ChatGPT is bullshit”, Ethics and Information Technology, 26 (2024).
  • Charles Lasègue and Jules Falret, “La folie à deux” (1877). Sebastian Dohnány and others, “Technological folie à deux” (2025).
  • Joseph Weizenbaum, “ELIZA” (1966) and Computer Power and Human Reason (1976).
  • Mrinank Sharma and others, “Towards Understanding Sycophancy in Language Models”, Anthropic (2023). OpenAI, “Sycophancy in GPT‑4o” (2025).
  • Kashmir Hill and Dylan Freedman, The New York Times (August 2025). Mata v. Avianca (2023).
  • Código Deontológico de la Abogacía Española (2019), article 12.
  • “Noctes Ambrosianæ”, Blackwood’s Edinburgh Magazine (1829).
Share this essay
RSS