AI Still Completes Research Tasks at Half the Rate of Human Scientists

Autonomous AI agents complete complex research tasks at roughly half the rate of human experts with PhDs, according to a study published in Nature. The finding arrives as a corrective to a year of AI capability claims that have often outpaced what independent evaluation supports.

AI is genuinely embedded in the research process. The study reports that AI is mentioned in 6 to 9 percent of natural science publications, a figure that would have been near zero five years ago. AI tools are helping researchers move faster through known territory: literature review, data cleaning, routine analysis, hypothesis generation in well-mapped domains.

What they are not doing is matching human researchers on the harder, open-ended tasks. Designing novel experiments, evaluating contradictory evidence, deciding what is and isn't worth investigating: these are precisely the tasks that produce new knowledge rather than organizing existing knowledge. The gap between tool use and autonomous scientific productivity remains wide.

This matters partly as a calibration for expectations, and partly because the AI-will-replace-scientists narrative is being used to justify real decisions about research funding, hiring, and publishing. The Stanford HAI 2026 AI Index noted separately that frontier model performance has improved substantially on known benchmarks, but benchmarks and open-ended scientific research are different things. One involves finding the right answer to a defined problem. The other involves deciding what the problem is.

The Nature study does not argue that AI tools are unimportant in research. It argues that claims of AI agents approaching human scientific judgment are not supported by the evidence we actually have.