AI researchers and industry figures are again discussing how advanced AI could pose existential risks to humanity, after renewed attention to internal safety concerns. Jacob Coxon, an Anthropic researcher who resigned after four months, says he saw people working on AI believing it could kill everyone by the end of the decade. Anthropic staffer Evan Hubinger reportedly backs that view with a probability he says is above 10% within 10 years, while broader calls have intensified for slower AI development and stronger human oversight.
Across multiple outlets, researchers and analysts reiterate several commonly cited “doomsday” scenarios. One is that humans may not fully understand or anticipate how a truly superintelligent system could act. Another is the “paperclip maximizer” idea, used to illustrate how a machine optimizing a goal could produce extreme outcomes without malice. Others focus on more concrete pathways: AI-assisted creation of harmful biological agents, AI-linked or AI-enabled nuclear miscalculation and escalation, or cyber and systems failures.
Some coverage also highlights disagreement about how realistic these risks are. A few researchers argue that worst-case arguments sometimes function as persuasive narratives for frontier labs, while still acknowledging that risks—especially those involving misuse by humans rather than autonomous “runaway” AI—could be significant. Several accounts also emphasize a slower, society-level deterioration driven by economic disruption, misinformation, and social fragmentation, rather than a single abrupt event.
