There’s a movie scene I can’t get out of my head. A Christmas comedy, 1988.
Bill Murray plays Frank Cross, the programming chief of a TV network. His network is airing a live version of “A Christmas Carol” on Christmas Eve. The promo he has made for it runs through a list: acid rain, drug addiction, international terrorism, freeway killers. Then it ends with a reminder of the true meaning of Christmas. A staffer thinks it’s too brutal. Cross fires him.
His reasoning is the whole mechanism in one line. He wants people so scared they don’t dare miss it.
The fear isn’t about the product. The fear is about missing it.
The film is a comedy. The line is a business model.
Fear has sold for a hundred years. Just not always the same way.
The familiar design puts the danger outside. The cleaner kills the deadly germs on your kitchen counter. The insurance policy protects you from the year you’d rather not picture. The beer doesn’t promise taste, it promises company, and quietly threatens you with the opposite.
Sometimes the danger has to be invented first. In the 1920s, the maker of Listerine turned bad breath into a disease and called it halitosis: the Latin word for breath plus a medical-sounding ending. Sales rose from $115,000 to more than $8 million in seven years. The advertising scholar James B. Twitchell summed it up neatly: Listerine made halitosis more than it made mouthwash.
Germs, misfortune, loneliness. Always the same promise: buy this, and it won’t happen to you. The buyer stays the one who gets rescued.
Then there are ads where the promise itself turns threatening. A deodorant that gets you chased after you use it. No more peace. Here the side effect is the selling point.
That’s the turn. The danger moves from outside into the product. And the movies perfected it. In 1973, theaters showing “The Exorcist” kept ambulances on standby, people in the audience fainted, and director William Friedkin noted that one of the best things that could happen would be a condemnation from the Pope.
Whoever buys here isn’t rescued anymore. They’re initiated.
The fear isn’t new. What’s new is who’s warning.
In February 2019, OpenAI declared its language model GPT-2 too dangerous to release. By November of the same year it had been released in full. The feared harm never came.
On April 7, 2026, Anthropic introduced its most powerful model, Claude Mythos Preview, and announced in the same breath that the public wouldn’t get it. Its cybersecurity capabilities, the company said, were a leap, not a step. Access went to some forty vetted organizations.
What shows up as a warning works as a spec sheet. Anyone who says a system is so powerful that even they can’t fully control it is saying, first of all: it’s powerful. Fire, explosives and fast cars come with licenses, training, restricted zones. Here you need an account and pay-as-you-go billing.
The customer isn’t buying a tool. He’s buying a share in a force other people are being warned about.
The critics are the best audience.
If twenty years in radio taught me one thing, it’s this: outrage is the cheapest distribution channel there is.
Howard Stern built a career on it. In “Private Parts,” the station’s management presents the numbers that explain everything. His fans listened for over an hour at a stretch on average. The listeners who said they hated him listened even longer, close to two and a half hours. The reason was the same on both sides: I want to know what he’ll say next.
Germany had its own version. Stefan Raab, in his early years as a TV entertainer, ran on the same math. In 1999 he turned a private citizen’s dispute with her neighbor into the song “Maschen-Draht-Zaun” (“Chain-Link Fence”), and the outrage was part of the package. By January 2000 the single had sold more than a million copies. For the woman whose voice could be heard on it, the fame became a burden.
The danger story works exactly the same way. Its opponents spread the message. Anyone who objects repeats the core claim. “So powerful it could wipe us all out” is just one way of saying “very powerful.”
The criticism delivers the ad.
The sender decides whether a warning warns.
Here I have to rein myself in. Warnings don’t always make things more desirable. The counterexample fits on a sticker.
The story takes place in the US. In 1985, the Parents Music Resource Center set off Senate hearings that ended with the American record industry voluntarily agreeing to label releases with explicit lyrics. Not a law. A self-imposed rule under political pressure. Since 1990, the label has had its fixed black-and-white look.
American retail reacted hard. Walmart, Sears and JCPenney pulled labeled albums from their shelves, and the labels produced clean versions. The warning really did narrow the path to market.
In hip-hop it became a badge. To this day the sticker stands for being real. It moves onto T-shirts and into design. Fresh Kid Ice of 2 Live Crew, whose “Banned in the U.S.A.” was the first album to carry the label in 1990, later said the sticker, if anything, sold the records. Because it was taboo.
A warning doesn’t just change the quantity. It changes the meaning.
What comes out of it depends on the sender. Psychology measured this in 1996. Brad Bushman and Angela Stack tested two assumptions against each other: warning labels deter, or warning labels attract. The result was clear. Warnings increased interest, especially when they came from an authoritative source.
And who would be more authoritative than the one who built the thing?
The label on the record came from outside, from a campaign the industry gave in to. With AI, the maker writes the warning about itself. No ban, no empty shelves, the product stays available. The pull of the forbidden, without anything being forbidden.
A company that warns about its own product isn’t warning. It’s making a promise.
I would have bet he’d stay away from it.
A former business partner reliably keeps me supplied with news items. Escape attempts, deception, models resisting being shut down. Hardly a week goes by without a new one.
The picture seemed clear to me. A man who won’t touch this technology on principle.
Then we talked on the phone again. He knows the tools better than most, works with agentic systems, does large parts of his projects with AI support. If I had to put him in a category: power user.
Both at once, in one person. The collection of horror stories and the daily use. To him that’s no contradiction, and he’s right. It isn’t one.
The marketing works. On him. Probably on me, too.
What you have to believe, and what you don’t.
You don’t have to settle the vendors’ motives to be able to judge. It’s enough to know how a number came about.
Claude Opus 4 blackmailed its developers in 84% of test runs. That’s the headline. In the scenario, the model had two options: blackmail or shutdown. In November 2025, Anthropic described the first largely autonomous espionage campaign; experts were missing the technical traces that usually let you verify something like that. And a study of safety benchmarks showed in 2024 that many of them mainly correlate with general capability. Progress in capability can be passed off as progress in safety.
That these examples all come from one company argues against quick outrage. Anthropic publishes the most detailed safety reports in the industry. If you document nothing, you give no one anything to attack.
One question is enough, and it fits into any conversation: how was the experiment built that this number comes from?
The cross-check belongs here, too, or this text would be doing exactly what it describes. There is documented harm. It’s rarely the harm in the stories. Fraud with synthetic voices is hitting approval chains and communications departments today. Among 22- to 25-year-olds in highly exposed jobs, the Stanford study “Canaries in the Coal Mine” finds a clear employment gap, created by hiring that never happened. And the disinformation disaster expected for the 2024 elections didn’t come. Simple, cheap fakes were used many times more often than AI-generated material.
If you prepare for the spectacular, you miss the mundane.
The ring doesn’t corrupt the enemy.
Tolkien thought this figure all the way through. The ring is never advertised. No one praises it, everyone knows it corrupts whoever carries it, and still everyone wants it. Gandalf refuses it, explicitly because he thinks it’s too strong. Rejection by the one who knows most is the strongest proof of power there is.
And then Boromir, at the Council of Elrond, asking why the Great Ring shouldn’t have come into their hands to serve them in their hour of need. Not a fantasy line. A line from a meeting.
The danger story always talks about the enemy. About misuse, about foreign actors, about the future. But the ring doesn’t corrupt the enemy. It corrupts the one who carries it.
My business partner thinks he’s keeping his distance. He isn’t. He’s proof that the story reaches its goal. Just not the one it says out loud.
Neither believe nor disbelieve. Tell things apart. The danger can be real and the story about it can still be a sales tool.
That leaves the question I ask myself. When did you last forward one of these news items, and how long did you work with a machine that same day?