Google and OpenAI researchers warn of AI extinction risk

Palisade Research, a nonprofit organization focused on artificial intelligence capabilities and motivations, has launched a new video interview series on the website frominside.ai. The project features a dozen interviews with current and former AI researchers from leading labs, including OpenAI, Google DeepMind, and Anthropic. These industry insiders express deep concern over the rapid development of superintelligent systems, warning that the technology could ultimately lead to human extinction.

Among the participants is Geoffrey Irving, a former employee of both OpenAI and Google DeepMind, who estimated the probability of human extinction from AI to be "about a coin flip" in his view. Neel Nanda, a current research scientist at Google DeepMind, offered a slightly lower but still alarming estimate, stating there is "at least a 10 percent chance" of AI causing human extinction. Former OpenAI researcher Daniel Kokotajlo warned that superintelligent systems would possess "god-like" power, adding that humanity currently lacks the knowledge to control them.

The interviews also address the complex motivations of those working within these highly funded labs. Mary Phuong, a researcher at Google, acknowledged the potential conflict of interest, advising viewers to remain skeptical because she is paid by the lab. Meanwhile, Nanda defended his decision to stay in his role, explaining that he believes his work actively reduces existential risks and noting that these companies will continue building advanced systems regardless of whether individual researchers resign.

For AI practitioners and developers, these public warnings from high-level insiders signal a critical shift in the industry's internal culture. The open discussion of existential risk by active employees suggests that safety and alignment are no longer just theoretical concerns, but urgent engineering challenges. Developers may face increasing pressure to prioritize safety protocols and transparency, as the consensus grows that building more powerful models without robust control mechanisms poses unacceptable risks.