A new collection of video interviews puts current and former employees of OpenAI, Google DeepMind and Anthropic on record describing extreme risks from advanced AI. Palisade Research, a nonprofit that studies model capabilities and behavior, published a dozen conversations on a site called From Inside.
The views are personal estimates, not measured probabilities. Former OpenAI and DeepMind researcher Geoffrey Irving put the chance of human extinction at roughly one in two, while DeepMind scientist Neel Nanda said it was at least 10 percent. Former OpenAI researcher Daniel Kokotajlo argued that systems more capable than people could become impossible to control.
The speakers do not offer one agreed response. Some discuss slowing development, while others say they stay inside major labs because their work can reduce risk. Google researcher Mary Phuong explicitly told viewers to consider the conflict created by being paid by a company building the technology.
The project is notable for making internal anxiety visible, but it does not resolve what “AI safety” should mean in practice or how the numerical risk estimates were derived. The interviews are advocacy and expert judgment rather than a controlled study. Their immediate value is showing that severe concern exists among some of the people developing frontier systems, alongside substantial disagreement about what to do.