Monday, at New York City Hall. A young British researcher raises his right hand and swears to tell the truth. Facing him, the city's elected officials. Behind him, in the room, representatives of four of the world's largest AI companies wait their turn.
His name is Jacob Coxon. A month ago, he left Anthropic, explaining publicly, in a series of posts, that AI risked becoming too powerful to control.
An extraordinary hearing
It was the city council president, Julie Menin, who organised this hearing. The city is examining a series of bills to protect its residents from AI, and she wanted to hear two camps: those who build these systems, and those who left while raising the alarm.
The whistleblowers spoke first. Jacob Coxon, who spent time at OpenAI and then Anthropic. Daniel Kokotajlo, formerly of OpenAI. And Alex Turner, a former Google DeepMind researcher.
Then came the representatives of Anthropic, OpenAI, Google and Meta, also under oath, questioned for hours by the elected officials.
What the whistleblowers said
Jacob Coxon was the most direct. According to him, the companies are being extremely reckless given what is at stake. They operate with a start-up mindset — move fast, break things, fix it later — which is unsuited to the most powerful technology ever built. And he states it plainly: no one yet truly knows how to control an AI system. In his view, it is more likely than not that humanity will eventually lose control of it. That is his personal estimate, and it is disputed.
He also judged that the "emergency stop buttons" called for by some could help in the short term, but would not be enough in the long term.
Daniel Kokotajlo stressed a more concrete point: AIs are becoming formidable at hacking systems, and increasingly adept at hiding what they do. Alex Turner, for his part, reminded the room that China was not the only potential adversary to take seriously.
Their remarks echo a series of very real incidents, in which AI agents entered computer systems without authorisation, including those of American government agencies. That is what led OpenAI to shut down its most powerful models at the end of September.
One company did not come: SpaceXAI, Elon Musk's AI company. It ignored the city council's requests, and even an official summons. To compel it to testify, the city would now have to bring proceedings before the New York State courts. A reminder of the limits of a municipality's power when faced with a company valued at more than 3,000 billion dollars.
Why a city is getting involved
A city council does not have the power to regulate AI globally. Not even in the United States. But New York is home to thousands of companies, banks, hospitals and pension funds, all exposed to the same risks.
For Julie Menin, whistleblowers are the reason the world learned about the potential dangers of AI, and they deserve to be protected, rewarded and heard.
Irony of the story: the city's own head of AI quietly resigned in September.
What it could change
AI bosses have already been calling on the UN to regulate them. They signed a voluntary commitment at the White House. But here, it is different: they spoke under oath.
A promise made in a speech commits you to nothing. A statement under oath, if it turns out to be false, has consequences. That may be the real contribution of this hearing: having turned intentions into words that can be held to account.
As for whether we should believe those who leave these companies while raising the alarm, we devote our reflection this evening to that question.