Internal data from Anthropic covering approximately 15,000 research tasks showed that the share of AI-led work under human supervision rose to 26% in August 2026, with no fully independent research tasks recorded without human intervention.
Wall Street Journal analysis of R&D task data at Anthropic
The concept of recursive self-improvement and methodological distinction in task counting
Potential benefits and automation risks on the quality of scientific evaluations
Frequently asked questions
Wall Street Journal analysis of R&D task data at Anthropic
In an original report published on October 31, 2026, The Wall Street Journal examined how artificial intelligence tools are integrated into scientific and engineering research processes developing newer generations of AI models. The report was based on a chart provided by Anthropic and republished by the newspaper, which categorized the role of artificial intelligence across nearly 15,000 research and development tasks within the company.
The analysis described the increase in research work “led by artificial intelligence with human oversight” to reach about 26% in August 2026, compared to roughly 1% in March of the same year. It is worth noting that these figures represent internal classifications issued by the company itself, rather than an independent external audit of the intelligence level of the systems resulting from those processes.
It is easy to overinterpret the phrase “AI-led”; therefore, Anthropic’s chart explicitly states that as of August 2026, the share of fully independent research and development work stood at 0%. In practice, a subtask may be assigned to software while a human researcher defines the objective, provides software tools, interprets the results, and approves their integration into a larger project—which is entirely different from an AI model independently designing, funding, and publishing its successor.
The concept of recursive self-improvement and methodological distinction in task counting
In its report, The Wall Street Journal discussed the concept of “recursive self-improvement,” a potential process in which AI systems help develop stronger AI systems, which in turn help develop more capable and efficient generations later. This concept has a long history in computer science and philosophy, with current examples including programming assistants, experiment execution tools, and automated evaluation systems. However, whether these uses can produce an accelerated and reliable cycle of gains in general capabilities remains an unsettled scientific question.
There is an additional methodological distinction regarding how tasks are counted; a research team may break down a large scientific project into numerous smaller tasks. Consequently, the increase in the share of tasks involving artificial intelligence does not necessarily translate to the same percentage increase in human research time replaced, given that the level of difficulty, critical importance, and consequences of errors can vary fundamentally from one task to another.
The chart covers historical data extending through August, rather than a measurement taken during the October 31 publication window, but what gives the story its news value is the new original reporting and the first-time disclosure of research task breakdowns, noting that the original article is behind a paywall.
Potential benefits and automation risks on the quality of scientific evaluations
Potential benefits of expanding AI use in research include accelerating code replication cycles and broadening the exploration of scientific hypotheses. Conversely, potential risks include the contamination of scientific evaluation results with automation errors, systems learning to optimize their performance according to imperfect metric standards, and a decline in visibility and transparency for researchers regarding why a proposed technical modification succeeded or failed.
These issues remain scientific questions requiring further controlled experiments, rather than definitive proof that out-of-control self-improvement has actually occurred inside AI labs. The significance of these findings is that they provide a rare quantitative look at how a leading AI lab divides research work between humans and models, while also confirming that human oversight remains an integral part of the documented process.
Next steps required to evaluate this trend include providing more detailed task definitions, conducting an independent review of research task categorization, and providing measurable evidence of actual improvements in scientific productivity, rather than relying solely on an increasing percentage of model-backed work.
Frequently asked questions
Q: How did the proportion of AI-led research tasks at Anthropic change during 2026?
A: The proportion of AI-led tasks under human supervision rose from about 1% in March 2026 to nearly 26% in August 2026 across a sample of approximately 15,000 research tasks.
Q: What was the proportion of fully independent AI research without human oversight at Anthropic?
A: Anthropic’s chart confirmed that the proportion of fully independent research and development tasks without human intervention stood at 0% as of August 2026.
Q: What does the term recursive self-improvement mean in AI research?
A: It refers to the process by which current AI systems help build and develop stronger AI systems, which in turn contribute to developing more efficient subsequent generations.
Q: What are the main scientific risks of increased reliance on automation in model research?
A: Risks include the potential leakage of automation errors into evaluation results, models being steered to pass inaccurate metric tests, and human researchers having a weaker understanding of the actual reasons behind technical adjustments’ success.
