On a boat, a sailor jumps into the water shouting that the hull has been breached. The other passengers look at one another. Has he seen a leak that nobody else noticed? Or has he panicked?
From the deck, there is no way of knowing.
On Monday, in New York, three former researchers from the biggest AI laboratories testified under oath that their industry was moving too fast. One of them even considers it more likely than not that humanity will lose control of these systems. Should we believe them?
Reasons to listen to them
They have seen it from the inside. These researchers worked on the models, sat in the meetings, saw the test results before they were published. They know things that neither the public nor elected officials can know.
They are paying a price. Leaving an AI company in 2026 often means giving up very high salaries and shares worth a fortune. Speaking publicly exposes you to legal action, to being shut out of a very closed professional world. Such a cost does not prove that you are right, but it makes it unlikely that you are speaking lightly.
History has often proved them right. In France, it was a doctor, Irène Frachon, who raised the alarm about the dangers of Mediator, when the drug was being prescribed to millions of people. Many scandals, from asbestos to tobacco, have been exposed by people who refused to stay silent.
France has protected whistleblowers since the Sapin II law of 2016, strengthened in 2022. A person who reports in good faith a threat to the public interest is protected against reprisals from their employer, such as dismissal or a sanction. They can go directly to an external authority, without first going through their line management, and the Défenseur des droits can support them. The idea is simple: whoever takes a risk to warn others should not have to pay for it alone.
Reasons to be cautious
We should also look at the other side, without cynicism.
Nobody ever leaves without personal reasons. A disagreement with your superiors, a disappointment, a new ambition can colour the way you see the place you are leaving. That does not make the testimony false, but it is an invitation to cross-check it.
You only see part of the picture. Even from the inside, a researcher sees their team, their projects, their tests. They do not see everything.
Predictions cannot be verified like facts. Saying that an AI has hacked a system is a fact that can be verified. Saying that humanity has more than a one-in-two chance of losing control is an estimate, which nobody can prove or disprove today.
How, then, should we judge?
The right method is not to judge the person. It is to sort through what they say.
On one side, the reported facts: incidents, practices, decisions. Those can be verified. And in this particular case, many have already been confirmed publicly: AI agents have indeed entered systems without authorisation, and several laboratories have acknowledged it.
On the other, the predictions: what could happen, and with what probability. Those cannot be verified, they have to be weighed. You can find them excessive without dismissing the facts that accompany them.
The trap would be to reject the facts because you find the prediction exaggerated. Or, conversely, to swallow the prediction because the facts are accurate.
What we take away
Someone who slams the door is not necessarily right. Nor are they necessarily wrong.
But a house from which nobody can leave to say what they have seen is always more dangerous than a house where those who leave are listened to, even if what they say is checked afterwards.