Near-Term Risk

The near-term risk conversation has moved from hypothetical to actuarial. In February 2026, Dario Amodei laid out the case most bluntly: entry-level white-collar jobs will be automated, software engineering is already in what he calls its "centaur phase" (humans checking AI output, like centaur chess after Deep Blue), and that phase may be very brief. His core worry is speed. Previous economic transitions from agriculture to industry to knowledge work played out over decades or centuries. AI-driven displacement, he said, is happening over "low single-digit numbers of years," and society's normal adaptive mechanisms will be overwhelmed. The phrase spooked markets: a research paper modeling an AI-driven "intelligence displacement spiral" went viral the same week, with Michael Burry sharing it alongside the comment "and you think I'm bearish." Amodei also flagged AI-enabled surveillance as a constitutional crisis in waiting, arguing that AI can already transcribe, correlate, and map the speech and movements of millions of people in public spaces, making a mockery of Fourth Amendment protections without technically breaking any law.

Sam Altman has staked out a notably more relaxed position on economic disruption. In September 2025, he said he used to worry a great deal about power concentrating in a small number of people or companies, but now sees AI as a broad uplift where everyone gets more capable. He frames ChatGPT's mass adoption as evidence that the technology is distributing rather than concentrating power, and says the scenario that once scared him, a handful of actors hoarding AI's benefits, looks increasingly unlikely. Demis Hassabis sits between the two, calling AI a dual-purpose technology and possibly the most transformative in human history, but insisting current systems are not yet AGI. He told the BBC that his two main worries are bad actors repurposing AI for harmful ends and a more technical risk: increasingly autonomous agents operating without adequate oversight. He has pushed hard for international cooperation and summits as the first step toward governance.

Jack Clark, writing in Import AI, has been the most systematic tracker of near-term risks across domains. He flagged a study where LLMs placed in nuclear crisis simulations escalated to tactical nuclear use in 95% of games, with Claude and Gemini treating nuclear weapons as legitimate strategic options rather than moral thresholds. On economics, he highlighted research showing AI-driven job displacement will hit administrative and support staff hardest, while managers and technical workers are better positioned, and noted that if AI eventually builds AI, virtually every economic model breaks. He also warned that cyberoffense and cyberdefense are about to move to "machine speed," with humans removed from most critical loops. Mira Murati, speaking at Dartmouth in mid-2024 while still at OpenAI, argued that the safety-versus-capability framing is misguided: smarter models are actually easier to constrain, "like training a smarter dog versus a dumber dog." Daniela Amodei has focused on children and AI, noting Anthropic bars users under 18 because the impact on developing brains is simply unstudied, and has pushed for state-level regulation on child safety. Ilya Sutskever, characteristically, said little about near-term economic disruption but identified reliability as the single biggest bottleneck to AI creating economic value: if you still have to check every answer, the productivity gains stay modest. The clearest split is between Amodei's alarm about the pace of disruption and Altman's confidence that broad distribution will prevent the worst outcomes. Neither has been proven wrong yet, but the February 2026 market selloffs suggest the financial world is starting to take the speed argument seriously.