Study finds AI coding-agent use is sharply uneven across social science

AI use in research is broad, but coding-agent adoption is not

A new Anthropic study highlighted in The Decoder suggests that social scientists are not adopting AI coding agents evenly. While general AI use appears relatively balanced across groups, the use of coding agents such as tools that automatically generate program code is much more uneven, with sharp gaps by gender, discipline, career stage, and university rank.

The most striking result in the supplied material is the gender gap. Researchers with typically male names were reported to use coding agents more than twice as often as researchers with typically female names. The study says that difference persists even within the same disciplines and career levels, suggesting the divide cannot be explained simply by field composition alone.

Economists lead, education researchers lag

The disciplinary spread is also significant. Economists were reported as the heaviest users of coding agents, with adoption at 39%, while education researchers sat at the bottom at just 4%. That range points to a major structural divide in how different branches of social science are integrating AI into day-to-day work.

The dominant use case was code generation for data analysis, cited at 97% among coding-agent users. Only about a third used AI for writing text, according to the source material. That detail matters because it distinguishes coding agents from general-purpose chat tools. In this study, the key shift is not simply that researchers are using AI more. It is that some are beginning to rely on AI much more heavily in computational workflows than others.

Career stage and institutional rank shape adoption

The study also found that PhD students and postdoctoral researchers use coding AI far more than professors do, and that researchers at top-25 universities adopt the tools 40% more often than peers elsewhere. Those findings fit a recognizable pattern in technology diffusion: newer entrants and more resource-rich institutions often move faster when tools promise productivity gains.

But the pattern also raises harder questions. If coding agents make it easier to process data, prototype analysis, or accelerate paper production, unequal uptake could deepen existing academic hierarchies. Researchers with better access, stronger quantitative traditions, or more permissive local norms may widen their lead over peers who are slower to adopt or more skeptical.

Researchers expect personal gains but worry about field-wide effects

One of the most revealing tensions in the study is how respondents view AI’s impact on themselves versus their discipline. The supplied text says 88% rated AI’s effect on their own paper output above 5 on a 10-point scale, and half rated it 8 or higher. Coding-agent users were even more optimistic than other respondents.

Yet 70% were more upbeat about their own productivity than about AI’s broader impact on the social sciences. The authors suspect that researchers worry increased paper output could overwhelm peer review, intensify competition for attention, and worsen existing problems such as selective reporting and risk-averse incremental work.

That split matters because it captures a familiar technology dynamic: the tool looks beneficial at the individual level while appearing destabilizing at the system level. In academia, where incentives already reward output, speed, and visibility, even modest productivity improvements can have outsized institutional effects.

The larger question is who gets left behind

The study does not argue that coding agents are inherently harmful. If anything, it shows that many users see them as practical accelerators for research work. But the uneven adoption pattern suggests AI is not entering the social sciences as a neutral layer that benefits everyone equally.

Instead, it may be amplifying existing divides around technical skill, institutional prestige, and access to computational practice. The gender disparity is especially consequential because it appears wider for coding agents than for general AI use. If that pattern persists, one of the most important AI shifts inside academia could reproduce inequities rather than reduce them.

Why the result matters beyond social science

The paper’s immediate focus is research behavior, but its implications are broader. Coding agents are increasingly marketed as universal productivity tools. Studies like this suggest their uptake may depend much more on local culture, discipline norms, and prior technical confidence than the marketing implies.

That makes the findings relevant well beyond universities. If AI coding tools spread unevenly even among highly educated knowledge workers, organizations should expect similar adoption gaps in industry, government, and non-profits. The challenge is no longer just building capable tools. It is understanding who adopts them, who benefits first, and which existing inequalities they quietly intensify.

This article is based on reporting by The Decoder. Read the original article.