AI coding agents have arrived in empirical research, but adoption is slower and more uneven than the industry's promotional calendar would suggest. A survey of 1,260 social scientists conducted by Anthropic researchers in February and March 2026 found that while 81% of respondents have used AI chatbots in their work, only 20% have brought coding agents into their research workflows. The paper, published May 22, 2026, offers the first rigorous cross-disciplinary count of where tools like Claude Code and Codex actually sit in the academic pipeline.
The gap matters because coding agents and chatbots represent fundamentally different kinds of assistance. A chatbot helps a researcher write or edit code, one piece at a time. A coding agent takes a research question, a dataset, and a set of instructions, then writes and runs an analysis, interprets the output, and iterates, automating what had been irreducibly human steps in empirical work. The researchers behind the paper argue that agentic coding platforms are, for the first time, capable of running the kind of multi-step empirical workflows that define quantitative social science. The question is whether social scientists are using them that way.
What the Survey Found
The 1,260 respondents were drawn from economics, political science, sociology, and adjacent fields. Recruitment targeted active researchers across career stages, from doctoral students to tenured faculty, at institutions in North America, Europe, and parts of Asia. The February-March 2026 fieldwork window captured a moment roughly eighteen months into the broad commercial availability of coding agents as standalone products.
The 81% chatbot adoption figure is high but consistent with other surveys of knowledge workers from the same period. What the social science data adds is a field-specific breakdown: researchers most commonly used chatbots for writing and editing code (cited by the majority of AI users), prose editing, and literature searching. These are tasks where the output is discrete and easy to check. The 20% figure for coding agent adoption is more striking because agents ask for a different kind of trust: you are handing the tool a task, not a line of code, and reviewing what comes back.
Social Scientists and Coding Agents: By the Numbers
- Survey respondents1,260
- AI chatbot adoption rate81%
- Coding agent adoption rate20%
- Gender gap in coding agent use2x (male vs. female researchers)
- Top-university adoption premium40% more likely than peers
- Fieldwork datesFebruary-March 2026
Who Is and Isn't Adopting
Two demographic gaps stand out in the data. Researchers with typically male names use coding agents at roughly twice the rate of those with typically female names. The paper is careful about causality: the survey cannot determine whether this reflects differences in programming background, access to training, confidence, workflow fit, or some combination. But the gap is large enough that the authors flag it as a concern for the technology's impact on research equity.
The institutional gap is also significant. Researchers at top-ranked universities are about 40% more likely to use coding agents than colleagues at other institutions in the same discipline and career stage. This follows a pattern seen in other technology adoption cycles: early adopters cluster at resource-rich institutions, which then accumulate productivity advantages that compound over time. The concern for the research community is that coding agents may accelerate the stratification of academic output in ways that have nothing to do with the quality of the underlying ideas.
The productivity signal, where it exists, is real. Researchers who have adopted coding agents are posting more working papers and applying for more grants relative to peers at similar career stages. They are also starting more projects. As of March 2026, they are not yet driving a surge in journal submissions, which the authors suggest reflects the time lag between project initiation and the publication pipeline rather than any ceiling on the tools' usefulness.
"Agentic coding platforms can take a research idea and a dataset, write and run an analysis, interpret the output, and iterate autonomously, automating what had been irreducibly human steps in empirical research for the first time." Anthropic, Coding Agents in the Social Sciences, May 22, 2026
The Trust and Verification Problem
The 80-to-20 gap between chatbot and coding agent adoption points to something more than a learning curve. Coding agents require researchers to trust outputs they have not produced and may not be equipped to fully verify. A social scientist who uses an AI to help write a function can read the function and check whether it does what it says. A social scientist who hands an agent a replication task and receives a completed analysis faces a harder auditing problem: the agent may have made choices about data cleaning, specification, and interpretation that are not immediately visible in the output.
This verification problem is discussed in the paper as a structural friction on adoption, distinct from skill or access barriers. Researchers who do use coding agents report spending significant time on output review, and several interviewees described abandoning agent-assisted pipelines on tasks where the review burden exceeded the time savings. The implication is that coding agents are more useful for researchers who already understand the underlying analysis well enough to spot deviations, which partially explains the institutional and gender gaps: both correlate with programming experience and research training depth.
The paper also notes that about 27% of the work coding agents actually do consists of tasks the researcher would not have undertaken at all without the tool. This "task expansion" effect is potentially significant for the research enterprise as a whole: it suggests that coding agents are not just making existing work faster, but enabling research that would otherwise stay undone. For fields like political science and economics, where large-scale data analysis has historically been bottlenecked by researcher-hours, that could accumulate into a meaningful expansion of what gets studied.
What Comes Next
The paper's most direct policy implication is that the benefits of coding agents are not distributing themselves evenly, and that absent deliberate intervention, the gap between adopters and non-adopters is likely to widen. The authors recommend that academic institutions invest in agent-specific training, that funding agencies consider adoption infrastructure as a research capacity issue, and that tool developers pay attention to the verification problem rather than optimizing purely for task completion.
For Anthropic, the research adds a dimension to the company's case for Claude Code and Claude Code as an enterprise platform: it is not just that individual developers are more productive, but that coding agents appear to be changing the structure of empirical inquiry in fields that have never been Anthropic's primary market. The company's research institute agenda explicitly includes studying AI's effects on knowledge production, and this survey is a data point in that longer project.
The sample is limited to social scientists, and the survey dates to early 2026, before the rollout of several features, including Claude Code's dynamic workflows and improved multi-agent orchestration, that could shift adoption rates. A follow-up survey in late 2026 will likely tell a different story. But the baseline it establishes, the 20% adoption figure, the 2x gender gap, the 40% institutional premium, gives researchers and policymakers something concrete to track as the technology matures. That is more than most AI adoption discussions have had to work with so far.
