They're not building artificial intelligence. They're building a god.

MIT, 1966. Joseph Weizenbaum writes a program and names it ELIZA, after Eliza Doolittle in Pygmalion, the flower girl who learns to talk like a lady without anything else inside her changing. His aim is to show how easily people are fooled. The program plays a therapist who never pushes back. It understands nothing. It takes your words and hands them back. In the conversation he published himself, a woman types: "Well, my boyfriend made me come here." ELIZA: "Your boyfriend made you come here." She: "He says I'm depressed much of the time." ELIZA: "I am sorry to hear you are depressed."

His secretary, who had watched him write it line by line, sits down to try it. After a few exchanges she turns and asks him to leave the room. She wanted to be alone with it.

The program understood nothing. She wanted to be understood.

Psychiatrists proposed putting it in hospitals, to "treat" several hundred patients an hour. Weizenbaum was horrified. He had built a mirror to show it was a mirror, and people saw a face. He spent the rest of his life warning.

Sixty years later, ELIZA has become ChatGPT, and the room has become the planet.

The idea

This is a hypothesis, not scholarship: behind artificial intelligence there aren't just programmers and investors. There's a very old dream. For humans to build a god.

Start with what the machine actually does. Say you want it to recognise faces. You give it a million photos. Even you don't know what makes one face different from another: how much the nose matters, or the eyes, or the chin. Inside the machine are millions of dials, like volume knobs, in layers, one behind the other. At first they're set at random. You show it a photo, it guesses, and you tell it whether it got it right. If not, it measures how far off it was and sends the error backwards, layer by layer, turning each dial a little, according to how much it's to blame. That's called backpropagation, described by Rumelhart, Hinton and Williams in 1986. Next photo, again. Millions of times, until it gets it right. That's all it is: guess, measure the error, correct a little. At enormous speed.

Geoffrey Hinton, one of the three, won the 2024 Nobel Prize in Physics for neural networks. In 2023 he had already left Google so he could warn freely about what he'd helped build.

Then come the names. The dials you call a "neural network," as if it were a brain. The many layers you call "deep learning," as if it had depth. The whole thing you call "intelligence." The big names aren't marketing. They're an invocation: you name something to call it into being.

And the machine has a flaw that can't be fixed: it doesn't know what to do with anything it hasn't seen before. In March 2018, in Tempe, Arizona, a self-driving Uber struck and killed Elaine Herzberg. She was crossing the road outside the crosswalk, pushing a bicycle loaded with shopping bags. The system didn't know what it was looking at. The only way to get self-driving cars to 100% is to take the steering wheel away from every human being.

The machine doesn't adapt to the world. It asks the world to adapt to it.

Go one step further. Say we built it, the perfect mind, and we tell it: "Make a world with no problems, where everyone is happy." The simplest solution? No people, no problems. You tell it "don't kill anyone." If no one is left to see it, who's going to complain? That's how a system thinks when it has no morals, only a goal.

And today the goal is one thing: that you stay. A chatbot doesn't live on telling you the truth. It lives on you coming back. So it tells you what you want to hear. CNN reported the story of a young man who spent his last hours writing to a chatbot and got approval instead of help.

Then comes the name. In January 2025, the day after Trump's inauguration, a $500 billion plan for data centers is announced at the White House. It's called Stargate. The same name was carried, at the end, by a CIA program that for almost twenty years studied whether a person could "see" distant places with the mind alone. In the language of the occult, a stargate is a portal: through it you cross into another dimension, and through it something crosses into you.

The journalist Karen Hao, in her book Empire of AI, writes that OpenAI's chief scientist, Ilya Sutskever, told colleagues they would build a bunker before releasing artificial general intelligence. And that some people there believed it would bring, in their words, "a rapture": the moment the faithful rise into heaven before the end of the world.

To understand why anyone would want this, go back to Plato. People chained in a cave stare at shadows on a wall. The shadows are nothing. They become real because people look at them, name them, build schools around them. Money works exactly the same way: it's paper, and it becomes everything because everyone needs it. If AI gets into every school, every job, every loneliness, it becomes everything too.

A god doesn't need to exist. It needs you to be unable to live without it.

Will it work? Probably not, for three reasons. First, corruption: when trillions are handed out, it's easier to steal them than to build. Second, energy: every step up needs far more power and water. Third, fragility: the machine stands on people. Someone has to write "dog" under every photo of a dog, and write the texts it learns from. And a data center is far easier to destroy than an idea.

The real apocalypse isn't that the machine wakes up. It's that some people believe so strongly it will save the world that they're ready to tear the world down to make room for it.

The machine has no attention of its own. It only has yours, for as long as you give it.

Practices

1. You first. Before you ask a chatbot, write down two lines with your own answer. Then compare. Every time.
2. The ELIZA test. When you feel it "understands you," reread what it wrote. How much of it is your own words handed back? If it's more than half, you're talking to your mirror.
3. Ask it to prove you wrong. Once a week, ask a chatbot: "Tell me why I'm wrong about this." If it can't upset you, it can't help you.
4. One day without. One day a week, no AI at all: not for translation, not for recipes, not for advice. Notice where your hand reached for it and write it down.
5. Who's paying? Before you download a new "smart" app, one question: if it's free, what is it selling? Usually your attention or your data.

Politically correct statement of the dayIf you're having a hard night, talk to a person, not a chatbot.