Steven Pinker: AI doom fears overblown, favors safety engineering

Pinker's open letter marks a public break with the extinction-focused wing of the AI safety discussion, which had been dominated by claims of a 20 percent chance of human annihilation. By admitting he underestimated how capable large language models would become and how recklessly AI companies would release their products, Pinker concedes ground on the technology's trajectory while disputing the likely outcome. He places the blame for that rushed release on people from the same community that warns about the risks, a tension he frames as ironic.

His alternative list of hazards — bioterrorism, cyberattacks, and unchecked AI agents — is narrower and more tangible than the paperclip maximizer thought experiment he criticizes for confusing intelligence with a drive to dominate. That distinction matters because the remedies he names, independent oversight, liability, and human control, are the kinds that can be designed and enforced now, unlike speculative extinction scenarios. Scott Alexander's own June position, meanwhile, supports better safety research and international agreements to slow development, so the two are not entirely at odds on the need for action.

The stakes here hinge on whether Pinker's framing can move the safety conversation from rhetorical warnings toward engineering and legal safeguards, and whether that shift gains purchase with the labs and policymakers who would have to implement it. Readers following AI governance may want to watch how the oversight and liability proposals he favors are taken up in practice, though the body does not specify any timeline or venue for that.