Anthropic warnings revive debate on AI extinction risks

This AFP piece crystallizes a striking disconnect at the heart of the AGI race: senior researchers inside leading labs are publicly saying they believe there is a non‑trivial chance their own work could wipe out humanity, yet the same labs continue to scale up and commercialize ever more capable systems. Jacob Coxon’s resignation from Anthropic, paired with Evan Hubinger’s blunt agreement that “we really do believe AI could kill all humans,” gives political actors and the public a simple, quotable narrative that will travel far beyond specialist circles.

For readers tracking the trajectory to AGI, this matters less as a fresh risk argument and more as an accelerant for regulatory momentum. When hundreds of AI workers and CEOs sign an open call for a coordinated slowdown, then high‑profile insiders quit over safety concerns, it becomes much easier for lawmakers to justify licensing regimes, capability caps, or security‑based deployment delays. At the same time, the article notes that both Anthropic and OpenAI have already delayed some model releases for national‑security reasons, suggesting that de facto gating by governments is starting before any comprehensive law is in place.

The net effect is to raise the political cost of aggressive scaling while sharpening the narrative divide: some will see these warnings as overdue honesty that mandates strict control, others as fear‑mongering that entrenches incumbents. Either way, frontier labs are being dragged into a more public contest over who gets to decide how fast we move toward AGI.