# AI Researchers Report Rising Extinction Risk Estimates
Nearly one in five leading AI researchers estimated an 18 percent probability of human extinction from artificial intelligence in 2024, according to a survey of more than 1,500 experts. The findings underscore a growing consensus among those building advanced AI systems that existential risks deserve serious attention.
The survey captured a critical moment in AI development. Researchers at Anthropic, OpenAI, and former Deepmind scientists have publicly articulated concerns about extinction scenarios. Anthropic researcher Jacob Coxon triggered substantial debate by flagging existential risks on social media. OpenAI researcher Daniel Selsam characterized the situation as a "ticking time bomb." These warnings come from people with direct access to frontier AI systems and intimate knowledge of their capabilities.
The 18 percent baseline reflects opinions gathered before the most recent wave of AI capability increases. The survey indicates that extinction probability estimates are climbing, suggesting researchers view current trajectory and near-term developments as moving the risk needle upward. This matters because these researchers operate inside organizations controlling trillion-dollar computational resources and foundational model development.
The gap between public perception and expert assessment remains stark. Most AI companies and media narratives emphasize near-term benefits like automation and productivity gains. Yet the researchers most intimately familiar with AI scaling laws, training procedures, and emerging capabilities assign meaningful probability to civilization-altering outcomes. An 18 percent extinction risk is not fringe thinking among this population.
Several factors drive these assessments. AI systems now exhibit unexpected capabilities that researchers did not explicitly train them to perform. The gap between predicted and actual performance has widened with each generation of models. Training procedures rely on human feedback signals that may not reliably capture human values or detect deceptive behavior in sufficiently advanced systems. As models approach or exceed human-level performance in reasoning and planning, alignment becomes technically harder while the stakes grow exponentially.
The timing matters. These estimates emerged in 2024, before GPT-o1 style reasoning models, before Deepseek's claims about new architectural efficiencies, before major breakthroughs in agent autonomy. Researchers surveyed were assessing risk based on systems and trends visible at that point. The fact that extinction probability estimates climb from there suggests researchers have observed developments that increase their concern.
This pattern echoes historical risk assessments in other fields. Nuclear physicists in the 1940s faced genuine uncertainty about whether atomic bombs could ignite the atmosphere. Climate researchers spent decades publishing increasingly dire forecasts as data accumulated. The consensus shifted not because experts were alarmist but because evidence pointed toward outcomes worse than initial models suggested.
The survey data reveals something important about expert disagreement on AI risk. Some researchers assign single-digit extinction probabilities. Others place the figure above 50 percent. This wide range reflects genuine uncertainty about how AI systems will behave at higher levels of capability, how alignment techniques will scale, and whether safety measures can keep pace with capability increases. The average of 18 percent thus represents a middle ground between significant concern and skepticism.
The implications extend beyond academic discussion. Researchers holding these views make concrete decisions about career direction, resource allocation, and research priorities. Many of the people designing safety infrastructure for advanced AI systems believe they are working to prevent extinction. This belief shapes their willingness to invest in technical solutions, advocate for slowing deployment timelines, and pursue speculative but important research directions.
As capabilities continue advancing, future surveys will reveal whether expert risk assessments have stabilized, increased further, or shifted toward lower estimates. The trajectory of these numbers provides a real-time window into how those building frontier AI perceive their own work.