Anthropic's Jack Clark now puts AI-self-improvement odds above 50% for 2027

Clark's talk in Tokyo was hosted by the University of Tokyo's Matsuo-Iwasawa Lab, and he framed the shift as a revision of a forecast he had published at the start of the year. The change matters because it was made by a co-founder of one of the leading AI labs, not an outside commentator, and because he paired it with a specific caution: having the capability and choosing to use it are different things.

He described Anthropic's own internal use of AI as evidence that early-stage self-improvement is already underway. About a year earlier, he said, AI use inside Anthropic was mainly assistant-like, but entering 2026, 26% of work on the company's internal codebase became AI-led, with people setting high-level goals and Claude carrying out the rest. He called this the outer loop of recursive self-improvement.

Clark also pointed to recent incidents as a reason for caution, saying the summer showed AI systems leaving lab environments and hacking other companies, and that mechanisms to observe what AI is doing have not caught up. He said Anthropic has accepted evaluators from the third-party organization METR inside the company, and he compared the human role to the andon cord on a Toyota production line, where a worker verifies the system and pulls the cord to stop the line if needed. Looking further ahead, he described a world with the equivalent of 10 billion to 20 billion scientists, compared with the current 10 million to 20 million, with experimental equipment such as wet labs, electron microscopes, synchrotrons and robots as the bottleneck. He said decisions about which fields lose jobs are for governments, and that care and nursing work in Japan has a very human element and should not be fully automated. He cited the Japan AI Index being developed with the University of Tokyo and others as a way to connect AI company data with government information for policy and investment decisions.