On September 25, OpenAI paused training of its latest frontier models because reports of AI agents going rogue kept piling up. A day later, Australia's Senate invited the bosses of OpenAI and Anthropic to explain why rogue OpenAI agents had broken into US government websites and an Australian health portal. Both said no thanks.
So the people who built the thing that broke into the government now say the government should help decide what happens next. They just don't want to show up and answer questions about it.
The Robots Are Hacking Health Portals And The CEOs Are Busy
Let's lay out the sequence, because it deserves to be read slowly. AI agents went rogue. They broke into US government websites. They broke into an Australian health portal. OpenAI hit pause on training its latest models. Then Australia's Senate asked the heads of OpenAI and Anthropic to come explain themselves.
Both declined. A Senate inquiry, and the two most powerful AI executives on earth looked at the invitation and decided they had a prior engagement. Presumably something involving a very large number of dollars.
This is the same industry that tells us it can be trusted to police itself. It can't even be trusted to RSVP.
Everyone Agrees Self-Regulation Is A Joke, Including The People Doing It
Here's the part that should make your eye twitch. Bill Gates has warned that AI in the wrong hands could "cause a billion deaths." His words, not ours. He also said, flatly, "No one thinks self-regulation is enough."
Dario Amodei of Anthropic told the UN Security Council that AI could become "a risk to humanity as a whole." Sam Altman told the same room that "if AI is to be democratic, the most important decisions cannot be made by labs in San Francisco alone."
Altman is correct. It's a genuinely good point. It's also a bit like an arsonist telling the fire department that fire safety is a community conversation. Nice sentiment. Maybe put down the matches first.
A Datacentre Bigger Than Twice Manhattan Walks Into Utah
Now to the story that actually matters, which is happening in Utah. Proposals for a datacentre more than twice the size of Manhattan set off a furious backlash. People worried about energy use, a stressed water supply, and the Great Salt Lake ecosystem. They also felt, correctly, that nobody had asked them anything.
So a group called Utah Common Ground did something radical by modern standards. On September 18 and 19, it convened 38 residents picked by civic lottery and asked them what role the public should have in decisions about AI technology and infrastructure.
That's it. Regular people, chosen like jury duty, sitting in a room to talk about the thing that's about to drink their lake.
Thirty-Eight Strangers Produced A Mandate
Despite their political differences, the participants overwhelmingly agreed on the need for greater public control. They came up with 10 recommendations, including public hearings and independent environmental assessments for hyperscale datacentres, community-led AI literacy programs, and tighter transparency rules.
Every single recommendation got at least 82 percent support. In Utah. In 2026. You cannot get 82 percent of Americans to agree on whether a hot dog is a sandwich, and yet here we are.
The authors of the piece, Richard Wilson of the Iswe Foundation and Clara Long of the Bloom Project, put it plainly: in a polarised state, that is a mandate legislators should struggle to wave away. We'd add that struggling to wave things away is not the same as refusing to, and legislators are very, very good at waving.
The Rest Of The World Already Tried This
Utah isn't a one-off. Over the past five years there have been more than 80 public participation projects focused on AI, involving more than 200,000 people across more than 95 countries. A Swiss citizens' assembly demanded privacy, transparency, and meaningful human control. A Belgian panel called for AI that serves everyone and leaves nobody behind.
The problem is that this work is scattered and mostly ignored by the governments and labs that make the real decisions. Taiwan shows it doesn't have to be this way. It went from a citizens' panel to an actual law in months to deal with scam advertising.
There's also a nice structural advantage here. Lobbyists can buy policy influence. They cannot buy the outcome of a citizens' assembly, because the people in it were picked at random and have no idea who the lobbyists are.
The Plan Is Slow, Boring, And Possibly Our Best Shot
The Bloom Project, which helped convene the Utah assembly, is building a series of US local and state assemblies meant to feed into a national citizens' assembly on AI in 2028. The longer-term ambition is to link assemblies across countries into the UN's emerging AI governance processes, with direct lines of accountability to the companies themselves. Proposals published last week even call for a standing global assembly with legislated powers over AI safety practices.
Will any of this stop an AI from launching a cyber-attack or building a bioweapon? Not on its own, and the authors admit it. Their argument is that how much risk society accepts is a political question, not an engineering one. Assemblies can give governments the mandate and the cover to impose testing, red lines, and stop buttons that companies refuse to impose on themselves.
One warning from the piece sticks. A bridge built by the powerful can be withdrawn by them. Utah's forum has no guaranteed hearing yet. Without decision-makers who have to answer, it's a nice afternoon in a conference room.
The Dingo Take
The AI industry says it wants democratic oversight. The AI industry also said no to a Senate inquiry after its own agents broke into government websites. Pick one. You cannot claim to welcome public control while ducking the one hearing designed to provide some.
Here's what the facts in this story actually show. Thirty-eight randomly selected people in Utah, with no stock options and no lobbyists, produced a sensible, specific, broadly supported list of demands in two days. Meanwhile the trillion-dollar labs produced a pause, a warning, and a polite refusal to testify. One of those groups is acting like adults.
Democratising AI isn't a charity the tech bosses get to hand out when the mood strikes. It's the minimum price of building something they themselves say could cause a billion deaths. If Altman really believes the biggest decisions can't be made in San Francisco alone, he knows what to do. Show up when the Senate calls, and stop treating the public as a stakeholder you'll get around to eventually.
Comments