Anthropic believes AI could soon play a much larger role in developing the next generation of AI models, potentially accelerating research that currently depends on human scientists.
In a Nov. 2 regulatory filing, the company warned that future systems could reach a threshold where their contributions dramatically speed up AI research and development. Such progress could shorten development cycles while making it harder for researchers to evaluate emerging risks.
The concern is not that Claude can already build a smarter successor on its own. Rather, Anthropic's research suggests AI agents are beginning to handle more of the experimental and engineering work involved in developing increasingly capable systems.
Claude is learning to conduct research experiments
In an April experiment, nine Claude AI agents investigated how a weaker AI model could help train a more powerful one. According to Anthropic's research, the agents developed and tested their own ideas, outperforming two human researchers on the experiment's scoring system.
Human researchers still chose the problem and determined how success would be measured. Although the agents performed well in testing, their best method failed to produce a statistically significant improvement on a production-scale model.
A separate analysis also found that current systems are better at carrying out experiments than deciding which research questions are worth investigating.
Faster AI development could outpace safety checks
Anthropic's Responsible Scaling Policy identifies a threshold for AI systems capable of dramatically accelerating research progress. One way to reach it would be for AI to help double the overall rate of AI capability improvements, a more demanding milestone than simply making researchers twice as productive.
In its amended SEC filing, the AI company acknowledged that it has previously failed to recognize significant acceleration in development until after it occurred. Some measurements become available only after progress has happened, potentially allowing more powerful models to advance faster than researchers can evaluate their risks.
Additional safeguards are under consideration, including mandatory human review and approval of certain research decisions as automated development expands. Existing precautions have included delaying model releases and restricting access to powerful systems, although the proposed approval measures have not been finalized.
What eWeek found: Claude is taking on more of its own development work
eWeek's review of a separate October engineering report found an early example of AI helping improve the systems used to develop it.
At Anthropic, daily merged code changes surged from roughly 2,500 in May to more than 30,000 in September, overwhelming existing infrastructure. A seven-person team then used Claude extensively to develop Scion, a custom version-control system that replaced Git across its main codebase within five months.
Another workforce study reveals how growing automation is also changing the responsibilities of AI developers. Employees increasingly direct agents and review their output, while junior workers accounted for a smaller share of new technical hires, falling from 27% in early 2025 to 9%. Overall employment continued growing, and the figures do not establish how much of the decline resulted from automation.
Scion provides a glimpse of how AI-assisted development could accelerate. Automated coding generated enough work to strain existing tools, and the same technology helped engineers build a system capable of handling the increased volume.
Together, the findings suggest that AI is already changing how research and engineering work gets done, even before more advanced forms of automated AI development emerge.
For companies adopting AI coding tools, the immediate challenge is ensuring that testing, code reviews, and engineering oversight keep pace with automated output. Preserving opportunities for junior developers to build technical judgment may become equally important as AI takes on more routine development work.
More AI news: Apple's latest AI benchmarks found that adding more Mac Studios doesn't always improve performance, with the results depending on how workloads are distributed.