AI Safety Needs More Evidence and Less Apocalypse

Summary

This opinion essay argues that artificial intelligence deserves serious scrutiny, but that claims of imminent human extinction require much stronger evidence and more nuance. It examines Dario Amodei’s repeated warnings that highly capable AI is arriving soon and could either transform science and medicine or cause catastrophic harm. The author notes that Anthropic is both building frontier AI and preparing for a possible initial public offering, creating incentives around a narrative that presents AI as extraordinarily powerful and urgent, while emphasizing that this does not prove Amodei’s concerns are insincere. The essay discusses former Anthropic researcher Jacob Coxon’s claim that there is a “strong chance we could all die,” but says employment at Anthropic does not establish knowledge of the future and that such a claim should be backed by observable evidence. It also notes that Anthropic’s leadership publicly supported Coxon and that his departure fits the company’s longstanding safety narrative, while acknowledging there is no evidence that the episode was orchestrated. The author agrees that AI risks are real and supports more independent university research, greater government technical expertise, information sharing about serious incidents, and meaningful access for independent evaluators. Those measures, the essay argues, do not depend on assigning a 10% or 20% probability to extinction. It calls for a prediction ledger so AI leaders’ claims can be revisited, and for concrete demonstrations of alleged capabilities such as autonomous vulnerability discovery, deception, or assistance with biological weapons. The broader case for caution rests on existing cybersecurity, biological misuse, labor disruption, and concentration risks, rather than on a confident doomsday forecast. The essay concludes that AI safety needs more measurement, reproducibility, and explicit conditions for being proven wrong, while noting that U.S. President Donald Trump opposed slowing AI development and said it would be much more beneficial than harmful.