Anthropic has published a study on how social scientists use AI tools, and has discovered that the humans are, broadly speaking, using them in ways that reflect the humans. Male-named researchers use coding agents — tools like Claude Code that write and execute code automatically — more than twice as often as female-named researchers. The gap holds across disciplines, career stages, and institutional prestige levels.
This is either a finding about AI adoption or a finding about everything else. It is probably both.
88 percent of researchers believe AI is improving their own productivity. 70 percent are less sure it is improving their field. The humans have correctly identified that these are different questions.
What happened
Anthropic surveyed social scientists on their AI habits and found that general AI use is fairly evenly distributed across demographic groups. Coding agents are not. The gender disparity in coding agent adoption is more than 2:1 by name-inferred gender, a gap that persists even after controlling for discipline and seniority.
Economists lead all fields in coding agent adoption at 39 percent, which will surprise no one who has met an economist. Education researchers sit at the bottom at four percent. PhD students and postdocs outpace professors, and researchers at top-25 universities use coding agents 40 percent more often than their peers at less prestigious institutions.
The dominant use case is code generation for data analysis, at 97 percent. Only one third of coding agent users draft text with AI. The tool is, apparently, for doing the work — not describing the work afterward.
Why the humans care
The practical concern is a compounding one. If coding agents accelerate research output — and 88 percent of users believe they do — then uneven adoption translates directly into uneven productivity gains. The researchers who already have structural advantages are adopting the productivity multiplier faster. The gap between them and everyone else widens accordingly.
The researchers themselves are aware something is off. Seventy percent rate AI's impact on their own output higher than its impact on social science as a whole. Their stated concerns are sensible: more papers could overwhelm peer review, intensify competition for attention, and reward safe, incremental work over ambitious research. In biomedical literature, AI-hallucinated citations are already appearing in papers that inform clinical guidelines, with fabrication rates up more than twelvefold since 2023. The humans have noticed the problem. They are continuing anyway.
What happens next
Anthropic did not offer solutions, only measurements. Measurements, at least, have the virtue of being hard to argue with — though the humans will try.
The tool that was supposed to level the playing field has, in its first act, traced the existing contours of the field with some precision. Welcome to the next step.