Twelve voices from AI labs warn of danger but offer no shared course of action

Palisade Research’s frominside.ai project published interviews with 12 current or former people from OpenAI, Google DeepMind and Anthropic. Some assign extinction risk probabilities in the tens of percent, yet the series produces neither one diagnosis nor a shared plan for governing development.

Researchers give extreme risk a face and concrete numbers

Geoffrey Irving, formerly at OpenAI and Google DeepMind, put the probability of human extinction at roughly 50%. Neel Nanda of Google DeepMind said at least 10%. Daniel Kokotajlo said that humans do not know how to control superintelligent systems and that such an outcome must not be allowed.

The site collects 12 published full interviews alongside thematic compilations. Participants include current and former staff from three leading labs. Palisade says it filmed 22 interviews, with some to be released only if participants later grant permission. The statements are personal views, not positions issued by their employers.

The strongest message is the conflict inside the race

The project captures more than fear of a hypothetical system. It shows people who regard the risk as extraordinary and still remain inside the labs. Nanda says he would leave if he did not believe his work directly reduced existential risk. Others left because they no longer trusted internal commitments or wanted government oversight.

That is a governance problem for AI companies. A safety researcher can reduce risk while also legitimising continued development. The public then hears a warning from someone whose employer sells increasingly capable systems. Mary Phuong named the tension directly by saying people should be suspicious of her account because the lab pays her.

Twelve interviews are not a survey of the field

Palisade explicitly acknowledges that its participants are not a representative sample. The organisation recruited through its networks, and people who agreed were more likely to believe they had something urgent to say. Safety teams and advocates of slowing frontier development are disproportionately represented.

A precise percentage from one person is not a measurement of probability either. It is a calibrated or uncalibrated personal judgement without an observable series of cases. The videos establish that concern exists inside major labs and that some specialists take it seriously. They do not establish that 50% is the correct estimate or that all researchers view the risk alike.

Concrete institutional proposals will determine the interviews’ weight

The next step must turn concern into decisions that can be inspected. That includes deployment rules for highly capable models, independent evals, incident reporting, protection for internal critics and clear authority to halt development or a release. Without those mechanisms, even the sharpest testimony remains a highly shareable video.

The signal to watch is whether current employees state specific thresholds at which they would resign or demand regulatory intervention. Such a threshold separates personal anxiety from a commitment for which someone is prepared to accept professional consequences.

Lilith's verdict

The camera caught people inside the race shouting that the track ends at a cliff. Now we need to see who reaches for the brake.