AI and coding agents are changing social science research | Keryc
The adoption of coding agents is already knocking on the door of social science. What does it mean for a machine to write and run analyses by itself? This article summarizes and explains, with data and common sense, what a survey of 1,260 quantitative social scientists reveals about the use of AI and coding agents in early 2026.
What are the "coding agents" and why do they matter
A coding agent is more than an assistant that suggests lines of code. Tools like Claude Code or Codex can take an idea, access a dataset, write and run analyses, interpret results, and iterate without constant human intervention. In technical terms, we’re talking about agentic platforms that automate empirical steps that used to be irreducibly manual.
Why does this change the game? Because automating those steps can speed up project timelines, lower the cost of experimentation, and potentially change what kinds of questions are worth investigating. But it also brings risks: congestion in the academic record, biases amplified by algorithmic choices, and inequalities in access to these tools.
What the survey showed: overview
Sample: 1,260 quantitative researchers surveyed in February and March 2026. It's not representative; respondents were recruited by offering access to Claude Max accounts, so the sample skews toward people interested in trying AI.
General AI use: 81% said they had used generative models to help their research (chatbots, assistants, etc.).
Adoption of coding agents: only 20% use a command-line integrated coding agent regularly (more than once a week). Claude Code is the most reported (86% among agent users), followed by Codex (31%).
In short: many people have tried AI, but few have integrated agents that write and run code on their own.
Disparities in adoption: who uses them and who doesn't
Adoption is clearly uneven. Some key figures:
By discipline: economists (39%) and political scientists (25%) use agents at much higher rates than researchers in public health (6%), education (4%), or communication (6%).
By career stage: PhD students and postdocs show higher adoption; tenured faculty adopt at less than half the rate of junior researchers.
By gender and status: researchers with typically male names use agents at more than double the rate of those with typically female names. Researchers at high-prestige universities use agents around 40% more.
These differences are statistically significant (p < 0.05) and suggest early diffusion favors those who already have more resources or greater pressure to be productive.
What they use AI for: code first, writing later
Contrary to the narrative that AI is replacing academic writing, the survey shows this:
97% of agent users and 77% of other AI users employ the technology to generate analysis code.
The second most common use is editing prose. Only about one third of all AI users have asked AI to draft text.
The practical takeaway: for now, AI is pushing empirical work and analysis engineering more than final article writing.
Does it increase real productivity? Early signals and limits
Descriptive comparisons show that agent users start more projects, publish more working papers, and submit more grant proposals than peers in the same discipline and career stage. Roughly speaking: they start about 0.25 more papers and publish about 0.5 more working papers in six months, after adjusting for discipline and stage.
But beware: this is not causal. Agent users might have been more productive beforehand. Also, there is no observed increase in journal submissions or faster resubmissions. It seems agents help accelerate the early phase and multiply ideas, but not necessarily push manuscripts all the way to final publication.
Researchers' expectations and concerns
Optimism about productivity: 88% rate above 5 on a 1–10 scale for whether AI will help produce publishable articles; the median hovers around 8.
Less optimism about effects on the discipline: 70% are more optimistic about individual productivity than about net impact on social science as a whole.
Why the gap? Researchers worry about congestion, lower quality, and automation reproducing bad practices (for example, selective reporting or weak replicability).
Important limitations of the study
Non-representative sample: respondents were recruited for access to tools, which biases toward those interested in or already using AI.
Self-selection and causal confusion: observed differences between users and non-users are descriptive. The authors are running a randomized experiment that will provide causal evidence in the future.
Quality versus quantity: the survey measures numbers of projects and outputs, not the quality or validity of those outputs. That remains an open question.
What to watch in the short term
Results from the randomized experiment that accompanies the survey; that will tell us whether giving access to Claude Code actually changes productivity causally.
Effects on inequality: if agents mainly help those who already have resources, visibility and funding gaps could widen.
Quality and peer review: if more working papers appear quickly, the review system may saturate and shift incentives toward publishing incremental over solid work.
Automation of empirical steps is real and expanding. For you—whether researcher or academic manager—the practical question is no longer whether you'll encounter coding agents, but how you'll integrate them so they improve science without worsening equity or quality.
