Intelligence Explosion Looms: Anthropic Warns AI May Self‑Upgrade by 2028

Anthropic’s research arm has just released a five‑page brief warning of an intelligence explosion — a scenario where AI systems accelerate their own improvement at a rate that outpaces human control.

Intelligence Explosion: Early Signs of Recursive Self‑Improvement

Co‑founder Jack Clark told reporters that the probability of an AI model training its successor by ↑ 60% before the end of 2028 is now “more likely than not.” He explained that today humans inject ideas into AI; tomorrow the machine may generate its own upgrades without external prompts.

“If you ask an AI to ‘make a better version of yourself,’ it could execute that command autonomously,” Clark said from Anthropic’s San Francisco headquarters.

The institute’s agenda clusters its work into four pillars: economic diffusion, threat resilience, AI governance in the wild, and the recursive self‑improvement question. It pledges monthly dashboards on labor impacts, and even proposes a “fire‑drill” tabletop exercise to test decision‑making across labs, boards, and governments.

Clark draws a parallel with Cold‑War hotlines, suggesting a global AI crisis channel could be as essential as the old nuclear link. “Rival nations will need a way to talk about existential technology,” he warned.

Anthropic also hints at coordinated policy levers, akin to central banks throttling inflation, that could modulate AI diffusion sector by sector. The move signals a shift from silent research to public accountability.

For further context, see coverage by Reuters and Bloomberg.

Intel provided by: Dr. Aris Thorne

Artificial Intelligence Researcher