Artificial intelligence is no longer only answering questions and writing code for users. At Anthropic, Claude is now helping to develop the next generation of artificial intelligence itself.
The company says Claude led 26 per cent of its artificial intelligence research and development work in August, a striking rise from less than 1 per cent in March. More than 90 per cent of the company’s research activity involved some form of collaboration between people and Claude.
Anthropic stresses that the system is not operating independently. Human researchers define goals, supervise work and remain responsible for decisions. Even so, the pace of change matters because an AI system that accelerates AI research can shorten the time between one generation of capability and the next.
Claude led 26 per cent of Anthropic’s research work, but the company says humans still set the goals and remain in control.
What “leading” research means
Leading a task does not mean Claude has become a scientist with its own agenda. It means the system can take a high level instruction, break work into steps, use internal tools and produce results that researchers evaluate.
Anthropic reported that about 30,000 AI agents were active on its internal platform in August. Their actions were screened before execution, and the company said only a very small share of decisions were blocked by safety systems. Those figures offer a rare glimpse into how a frontier AI laboratory is using its own technology behind the scenes.
The model can assist with experiments, software engineering, analysis and the search for improvements. That can make researchers more productive, but it also creates a loop: better models help humans build still better models, which may then contribute even more to the next cycle.
Why the disclosure matters
For years, recursive self improvement has been discussed as a future possibility. Anthropic’s figures do not show a fully autonomous system redesigning itself. They do show that the boundary between tool and research collaborator is moving.
That raises practical questions. How do researchers verify work produced at machine speed? What happens when thousands of agents interact? Can safety teams keep pace with capability teams? How much computing power is dedicated to safeguards, and who independently tests the results?
Anthropic says every agent action is screened and that human oversight remains central. It has also called for other laboratories to publish comparable measures so governments and the public can understand how quickly AI assisted research is advancing.
The human impact is already beginning
The immediate story is not a distant contest between people and machines. It is a change in how technical work is organised. Researchers may spend less time on routine coding and more time setting direction, checking evidence and judging risk. The skills that become more valuable may be the distinctly human ones: asking the right question, recognising weak reasoning, understanding consequences and accepting responsibility.
For countries outside the main AI development centres, transparency is essential. Decisions made inside a few laboratories can affect jobs, education, security and access to information around the world. Public understanding should not depend on occasional disclosures after capabilities have already changed.
A test for the industry
Anthropic’s report should begin a wider standard of disclosure. Leading AI companies should publish regular, independently reviewed information on how their systems contribute to research, what actions they can take, what failures occur and how human control is maintained.
Should AI laboratories be required to report when their models help build more powerful successors? Podium News invites researchers, policymakers, businesses and readers to share evidence based views on transparency, safety and accountability.
