Anthropic reports Claude leads 26% of its AI R&D tasks
On September 17, Anthropic proposed measures of AI development. Its August internal assessment put 26% of R&D at a Claude-led level under human oversight, using an index weighted by human time spent on tasks.
That level allows much of a task to proceed from a high-level prompt while people supervise. The company reported no measured research area at full autonomy.
Our analysis
Waiting for review could become the new research bottleneck
As AI produces more work, deciding what humans review first becomes more important. A hundred experiment proposals do not help equally if only a few people can examine them.
Competition could therefore shift toward identifying consequential failures and discoveries quickly. Labs that focus human expertise on the right checks may gain an advantage.
How could everyday life change?
The following is a possible future based on this news.
Choose the important overnight experiments each morning
A future researcher might review overnight AI work in the morning and pursue surprising results. With less routine execution, the questions people find puzzling could become more distinctive.
This scenario extends beyond the reported figure. It imagines a change in how people divide time between supervision and ideas, not the disappearance of oversight.
