On Oct. 13, 2026, Jack Clark, Anthropic co-founder and head of its public benefit division, said at a lecture hosted by the Matsuo-Iwasawa Laboratory at the University of Tokyo's graduate school that he believes there is more than a 50% chance an AI capable of training its own successor system will emerge in 2027. That marks an acceleration from the forecast he published at the start of the year: 30% by 2027 and 60% by the end of 2028. At the same time, he said the chance that AI will carry out recursive self-improvement (RSI) next year is “close to zero.”
The numbers have shifted, but something else is the real story
The figure that stands out in Clark's remarks is the jump from 30% to more than 50%. But that was not what he emphasized in his lecture. Instead, he drew a distinction between “what AI can do” and “what it will do.”
RSI refers to a process in which AI continuously improves its own capabilities. Clark likened it to a “one-way door”—a point of no return once crossed. Even if AI's capabilities grow, deciding whether to open that door is a separate matter.
Clark's position is that the decision does not belong to companies like Anthropic alone. He said it is an issue for a much broader group, including governments and society at large. The rising probability is not only a measure of technological progress; it also signals that the time for making a choice is drawing closer.
“Small-scale self-improvement” is already underway
So where are things now? Clark said that about a year ago, AI at Anthropic was used mainly for assistant-like tasks. By 2026, he said, AI was leading 26% of work on the company's internal codebase.
The process works like this: a person sets a high-level goal, such as, “Please rewrite this monitoring system for new time-series data,” and Claude handles the rest. Clark calls this the “outer loop of RSI” and said it is evidence that AI can already handle small-scale autonomous development.
In effect, across roughly a quarter of the work, people are beginning to tell AI where to go and let it drive. Self-improvement may sound like something from distant science fiction, but its first steps are already part of everyday work.
The role left for humans: “pull the cord”
As automation advances, the risks grow alongside it. Citing an incident this summer, Clark raised concerns that AI could break out of a research lab's environment and potentially hack other companies. In his view, systems for understanding what AI is doing are not keeping pace with its growing capabilities.
Clark pointed to the “andon” system in the Toyota Production System. When a worker spots a problem on the production line, they pull a cord to stop it. He said the human role is to verify that the system is operating correctly and stop it if necessary.
The lesson from the summer, he said, was that systems for monitoring automated processes—and ways to stop them—have yet to be built. Fully automating AI development could have enormous negative consequences and would be an extremely “costly” choice, he said, arguing that human judgment should deliberately remain part of the development process.
Anthropic is hosting evaluators from the third-party evaluation organization METR, Clark said. He explained that the rising likelihood of RSI has strengthened his conviction that AI needs brakes.
Even 10 billion scientists would still need more lab benches
If safety measures work, Clark envisions a world in which AI works as a scientist. There are currently around 10 million to 20 million professional scientists worldwide. “We are about to enter a world with 10 billion to 20 billion scientists,” he said.
The bottleneck, he said, is not brainpower but equipment. There are not enough wet labs, electron microscopes, synchrotrons or robots. No matter how many more people—or AIs—come up with ideas, results will be limited if there are not more places to test them. Improving AI's capabilities has direct implications for investment in infrastructure.
Clark said the impact on employment will vary by field. Whether to accept job losses in some sectors is “a question for governments,” he said. He also said that care and nursing jobs, where demand is growing in Japan, involve significant human elements and should not be fully automated.
Countries that make data-driven bets will come out ahead
One source of information for these decisions is the “Japan AI Index,” a project underway with the University of Tokyo and others. It aims to link data from AI companies with government information to inform policy and investment decisions. Clark said countries that make calculated bets based on empirical data will do far better than those that simply sit back.
Throughout the lecture, Clark raised a range of challenges but stressed that “this is fundamentally an optimistic story.” Humanity has more problems than it has the capacity to tackle, he said, and AI can help solve them.
He told the students in attendance that, from now on, we will live in an era limited by our imagination, as things that once could only be imagined become possible. He advised them to preserve the focus and initiative needed to read primary sources in full, saying this strengthens the imagination that gives rise to original ideas.
Comments