Where do researchers actually use AI in their work?

Nature Research Intelligence, in partnership with Fudan University, publishes "AI for Science 2026: The State of AI Use among Researchers" on nature.com (dated 2026-07), reporting that AI assistance "concentrates on information- and text-intensive tasks, but its role diminishes where judgment is required." Nearly half of respondents use AI "every time" or "most times" for information gathering (10.5% and 33.3%) and for improving or editing papers (12.2% and 32.4%), but fewer than a quarter report that same frequency for reviewing papers (5.8% and 17.7%). The survey frames AI as "reinforcing the exploratory and generative stages of research while leaving tasks that rest on accountability and professional judgment largely untouched."

The survey ties this pattern to seniority: frequent use for editing falls from about 46.8% among researchers with under three years' experience to 36.1% among those with 20 or more years, and from roughly 47.8% to 33.4% for information gathering. The authors attribute the decline to senior researchers opting out rather than to existing users scaling back, though the excerpt does not detail how they isolated opt-out from reduced intensity among continuing users. Open-ended responses reinforce the same task-selective pattern: across more than 7,500 comments, respondents described AI as augmenting rather than substituting for judgment, citing productivity and literature-discovery gains as the leading hopes, but naming accuracy and hallucination (4,935 mentions) as the single most common concern — more frequent than any hope cited.

This self-reported task boundary parallels How widely do peer reviewers actually use AI tools?, another publisher survey that finds the same early-career lean in adoption, though Frontiers counts any reviewer use of AI tools at all (53%) where this survey's "every time/most times" threshold for reviewing stays under a quarter — a reminder that adoption rate and frequency of use are not the same measurement. The judgment-task gap also echoes, from the systems side, What stops AI from discovering science without human help?: researchers withhold frequent AI use from the accountability-bearing stage of their own work, while agentic science systems are separately argued to lack the design features needed for the same judgment-heavy stages. Does AI help individual scientists while narrowing scientific focus? supplies a complementary, behavior-based measurement (citations and publication counts) against this survey's self-reported tool-use frequencies.

The excerpt gives no total sample size, response rate, or sampling frame, so how representative the pooled percentages are of researchers generally — beyond the country breakdowns given for Japan, China, the US, and Europe — is not stated. The seniority comparison is cross-sectional, across experience bands at one point in time, not a panel following the same researchers as they age, so "opting out" describes the pattern in the data rather than an observed behavior change in any individual. And because the survey is co-published by Nature Research Intelligence, Springer Nature's own research-analytics arm, as partner content, its framing of AI as augmentation rather than substitution should be read as the survey's own interpretation of self-reported frequencies, not as an independent measurement of actual judgment quality or capability limits.