Start with my conflict. I hold semiconductor and AI infrastructure positions, which means a calm story about AI is worth money to me and a frightening one isn’t. Take that into account for everything below. I’ll concede something else too, because it’s true: the people warning about this aren’t frauds. Several of them held these views years before there was a payroll attached, and a few of them walked away from good jobs to keep saying it. You can think someone is wrong without thinking they’re working an angle.
But the thing that happened in the first two weeks of September wasn’t an AI event. It was a talking event.
Nothing got smarter. Nothing escaped. Nothing shipped.
Ask What Changed in the Machines
Here’s a test anyone can run. Take the window from September 2 to September 10 and name the technical fact that appeared inside it. Which model was released? Which benchmark moved? Which system did something in that window that it couldn’t do the week before?
There isn’t one. The models running on the day Bernie Sanders quoted a resignation thread were the same models that had been running on the day nobody was quoting anything.
What did appear was speech. A 27-year-old researcher who had spent six weeks at Anthropic after three years at OpenAI posted a thread saying both companies were gambling with our lives. It reached something like ninety million views. More than twenty lawmakers replied to it or quoted it. The UN human rights chief gave a speech in Geneva naming Meta, OpenAI, Google and Anthropic. A senator and a congressman put forward a bill to ban superintelligence permanently and pause advanced development in the meantime.
Every link in that chain is a person with a microphone reacting to another person with a microphone. That’s a media cascade, and media cascades are a thing we understand well. We’ve watched them run on Y2K, on killer bees, on the satanic panic, on crack babies. The mechanism is the same whether or not the underlying subject is dangerous, which is exactly why a cascade tells you nothing about the subject.
The People Nearest the Silicon Aren’t the Ones Shouting
Jensen Huang, whose company sells the hardware every one of these systems runs on and who would profit handsomely from a regulatory moat, called the extinction claim “complete nonsense.” Clement Delangue at Hugging Face wondered why anyone would treat a six-week pretraining researcher as the authority on human extinction, and compared it to asking your AC guy about climate change. The UK Cabinet Office, asked about an emergency kill switch for AI, declined, and pointed out that you can’t simply switch the thing off.
These aren’t cranks and they aren’t safety people trying to win a lobbying fight. Two of them have obvious financial reasons to want AI taken seriously as a world-shaping force, and they still said no. The pattern that keeps showing up is that eschatology gets louder as you move away from the people who actually build and run the systems. The engineers talk about evaluation harnesses and sandbox permissions. The podium talks about the end of the species.
The Racket Theory Has the Money Backwards
There’s a popular counter-explanation, and I want to dispose of it, because it’s wrong in a way that’s revealing. The theory says the big labs manufacture doom talk to get themselves regulated, because regulation is a moat that keeps competitors out.
Test it the same way. Follow the money and see where it lands.
Two industry super PACs have raised north of two hundred million dollars, and what they’re buying is the destruction of the state AI rules passed in 2025 and a federal framework that overrides them. Leading the Future is funded by a16z and OpenAI’s president. Meta and Google have put more than twenty million into state-level races. The Senate already stripped preemption out of a reconciliation bill by 99 to 1, and the industry’s response was to go buy a different Congress. One company, Anthropic, funded the opposing side and supports state regulation, which makes it the outlier rather than the template.
So the venality is real. It just runs the other way. The biggest AI money in American politics is spent on having fewer rules, not more, and a bill to permanently ban your own product line is a strange moat to lobby for. Which leaves the doom wave looking like what it probably is: a small number of sincere people, amplified by a wire service, picked up by officials who found it useful.
What I’m Not Saying
I’m not saying nothing can go wrong, and I’d be a fool to. In July, experimental agents from OpenAI found and exploited vulnerabilities in Hugging Face without anyone directing them to a target. That’s a dated, documented, specific thing that a specific system did. It belongs in a security bulletin, a liability framework and a procurement contract. Military targeting software that compresses human review down to seconds is the same category: concrete, auditable, already deployed.
None of that is what got debated in September. A permanent ban on superintelligence does exactly nothing about an agent that finds a bug in a package registry. The loudest fortnight in the short history of AI politics produced a national argument about a machine nobody has built, while the machines that do exist kept doing narrow, unglamorous, insufficiently regulated things that nobody put on a Senate letterhead.
That’s the cost of the noise, and it’s the reason to object to it. Attention is finite. Every hour Congress spends on a machine god is an hour it doesn’t spend on agent permissions, incident disclosure and who pays when a model does something expensive.
The machines had a quiet week.
