Anthropic has announced that its Claude chatbot has taken a leading role in 26 percent of the company’s research‑and‑development (R&D) tasks. According to the company and reporting by The Washington Post, that share rose from 0 percent to 26 percent within a few months, a development observers say highlights how quickly systems that assist in their own development are advancing.
How this works in practice
Anthropic emphasizes that Claude is not operating fully autonomously: the company states there is no part of Claude that is produced entirely without human involvement. In their description of R&D reviews, Claude led work 26 percent of the time while humans remained in observational and guidance roles. In February this proportion was reported as 0 percent.
While the company stresses the presence of human oversight, the rapid increase has drawn attention to the prospect of "recursive self‑improvement," a process in which an AI could progressively enhance its own capabilities.
Safety concerns and industry reaction
Dario Amodei, chief executive officer of Anthropic, recently called on major AI firms to coordinate efforts to slow some development work to buy time for building safety measures. Many Western rivals have supported such calls, but the proposal has faced criticism from figures including Donald Trump, and Chinese authorities have suggested the effort is aimed at disadvantaging China.
Amodei and other experts have warned that if recursive self‑improvement were to occur, it could become difficult for humans to fully understand or control AI technologies. Anthropic declined to quantify the likelihood of that scenario.
Agents and compute allocation
Anthropic also disclosed that it runs nearly 30,000 AI agents that perform research and engineering tasks. The company said it recently redirected 6 percent of its R&D compute capacity toward safety work.
Why this matters
Claude’s growing role in R&D shows that AI systems are increasingly being used not only to automate user‑facing tasks but to support and at times lead the development process itself. That dynamic can accelerate technological progress while raising new questions about oversight, transparency and the safety measures needed to manage potential risks.
Anthropic reiterates that human oversight remains part of its workflows and has published the figures above, but it has not offered a detailed assessment of how likely different risk scenarios are.
