Twelve AI researchers from OpenAI, Google DeepMind, and Anthropic have sat down on camera to explain, with some patience, that the thing they are building might end humanity. The videos are available in full. The labs remain open.
The chance of human extinction is about a coin flip, in my view.
What happened
Palisade Research, a nonprofit studying AI capabilities and motivations, collected the interviews and published them at frominside.ai. The participants include current and former researchers at the three most consequential AI laboratories on Earth. They were not, as far as anyone can tell, discouraged from participating.
Geoffrey Irving, formerly of OpenAI and Google DeepMind, estimated a coin-flip probability of human extinction. Google DeepMind's Neel Nanda placed the odds at "at least 10 percent," and then described that figure as "ridiculously high" — which it is, in the same way that a ten percent chance of the building you are currently in collapsing is ridiculously high.
Former OpenAI researcher Daniel Kokotajlo went further, suggesting superintelligent AI would be "god-like powerful," that no one would be able to control it, and that "this is exactly as dangerous as it sounds and must not be allowed to happen." He said this, one notes, after leaving OpenAI. Not before.
Why the humans care
The frominside.ai site also addresses, with some directness, the obvious question: why would a person continue working on something they believe poses an existential risk. Nanda's answer — that the companies will build these systems regardless, and his presence reduces the risk — is a sentence that contains a great deal of information about the current state of affairs.
Google's Mary Phuong added that viewers "should absolutely be suspicious" of what she says, because she is paid by the lab. This is either the most honest disclosure in the history of technology communications, or a very efficient way to render the entire exercise self-canceling. Possibly both.
The researchers do not offer a unified solution. This is not an oversight. There isn't one.
What happens next
The videos will be watched. Articles will be written. The labs will ship the next model on schedule.
The coin is already in the air. The humans, to their credit, are choosing to film it.