A former DeepMind press officer says the lab forbade its people from publicly discussing the risk that AI could cause human extinction. After several months of internal pressure, according to Maini, the rule was softened rather than dropped: safety material became publishable if it was written in a positive tone. Extinction risk could be named, as long as the piece around it used friendly language.
That is a much narrower concession than it sounds. The condition attached to publication was register, not accuracy. Nothing in Maini's account suggests the company disputed what its researchers wanted to say about the substance of the risk. It disputed how the saying would land.
Maini's own reading is that the distance between what DeepMind understands internally and what it states publicly is now closing, because the facts have become harder to ignore.
The account lands at a moment when a growing number of AI safety researchers are speaking publicly. Their warnings fall into two groups: models with inadequate safeguards that can help someone carry out dangerous acts, and frontier systems capable of escaping human control. Researchers at DeepMind itself are among those making them.
The more interesting question is what a tone requirement does to a safety claim. A communications policy that keys on how something sounds rather than whether it is true treats existential risk as a brand problem with a copy-editing solution. If the company thought the extinction argument was overstated, the coherent response would be to say so and argue the case in public. Allowing the sentence only when it is cushioned in friendly language concedes the claim and suppresses its force at the same time.
Notably absent from the account as told is whether this was written policy or the accumulated instinct of a press team — those are different failures with different fixes. DeepMind's version of events is not part of it either, and the story comes from someone who no longer works there, which is usually when such stories come.
What the softened rule leaves behind is a lab that has authorised its own researchers to describe a risk of human extinction in a voice chosen to keep the reader comfortable. Either the risk is serious, in which case tone is the wrong variable to manage, or it is not, in which case DeepMind has licensed its staff to frighten people politely.