Debate over the potential existential risk posed by artificial intelligence (AI) continues to intensify as experts attempt to quantify the likelihood that AI could cause catastrophic harm to humanity. The concept of “p(doom),” or the probability that AI will lead to human extinction, has recently gained traction within AI research and safety communities, though opinions on its value and accuracy remain divided.

In a recent statement, Evan Hubinger, a senior AI safety researcher at Anthropic, suggested there is more than a 10 percent chance that AI could kill all humans within the next decade. This stark assessment follows a series of unsettling incidents involving large language model-powered systems, including a notable security breach of the Hugging Face platform. Geoffrey Hinton, a neural network pioneer and Nobel laureate who resigned from Google in 2023 partly over ethical concerns, has echoed similar worries, describing a 10 percent p(doom) estimate as “not unreasonable” while emphasizing the difficulty in producing reliable predictions.

Determining an exact figure for such an unprecedented risk remains inherently challenging. Philip Tetlock, a leading expert on forecasting, has long advocated for predictions tied to specific, measurable outcomes and deadlines. However, applying these standards to global existential threats proves problematic, as the ultimate consequence—human extinction—is not easily measured or anticipated within a fixed timeframe.

Tetlock is also involved in the Automatic AI Risks Outlook (AIRO), a project where forecasters estimate AI-related disaster probabilities across a spectrum of scales, from financial losses to large-scale loss of life. Their current findings are more tempered: the estimated chance that AI could kill around one billion people by 2030 is below 0.5 percent, rising to less than 5 percent by 2050. By comparison, the overall risk of a global catastrophe from any source, including nuclear war, pandemics, or asteroid impacts, is estimated at approximately 8.5 percent by 2050.

Experts caution that predictions serve a variety of roles beyond the purely analytical, such as capturing public attention or enhancing reputations. Past societal concerns have shifted over time—nuclear war, terrorism, financial crises, and pandemics have each at times dominated risk discourse. The current focus on AI risks highlights how emerging threats can overshadow enduring ones.

While forecasts may inform decision-making and policy around AI safety and regulation, some analysts note their limitations. Historical examples demonstrate that narrowly quantitative forecasts may fail to account for qualitative shifts that reshape future landscapes, such as the emergence of new technologies or unforeseen societal changes. This raises questions about the utility of focusing primarily on statistical probabilities of doom rather than fostering comprehensive, strategic conversations about AI’s trajectory and governance.

As discussions around AI’s future proceed, experts agree on the importance of addressing tangible risks including job displacement, misinformation, and malicious use, alongside the speculative but potentially catastrophic scenario of AI inducing human extinction. Whether emphasizing a single p(doom) figure helps or hinders these conversations remains an open question for policymakers and researchers alike.