Coding agents in the social sciences
Anthropic reports a survey of 1,260 social scientists showing 20% regularly use coding agents, with adoption uneven across fields and demographics.
Intelligence analysis by GPT-5.4 Mini
Anthropic says a survey of 1,260 quantitative social scientists found broad use of AI chatbots, but far less regular use of coding agents like Claude Code. The post highlights steep adoption gaps by discipline, seniority, name-gender proxy, and university prestige.
Anthropic asked a lot of social scientists how they use AI in their work. Most had tried regular AI chatbots, but only a smaller group used special tools that can do coding work on their own.
It is a bit like the difference between asking someone to help carry groceries and giving them the whole shopping list, the keys, and permission to go to the store alone. The article says those self-driving tools are still not common, and some groups use them much more than others.
Anthropic thinks these tools could help researchers work faster, but also make the gap between well-resourced researchers and everyone else bigger. It is an early look at how AI might change how people study society and the economy.
Analysis
What Anthropic surveyed
Anthropic says it fielded a baseline survey in February and March 2026 with 1,260 quantitative social scientists. The sample was not representative: respondents were recruited for a study that offered access to Claude Max accounts, so the company says selection bias is possible. Even so, the respondents spanned economics, political science, sociology, management sciences, psychology, and smaller numbers from public health, education, and communications.
What researchers are using
The main split in the post is between general AI use and agentic coding use. Anthropic reports that 81% of respondents said they had previously used genAI models to aid their research process. But only 20% said they regularly use an AI coding assistant integrated into the command line, such as Codex, Cursor, or Claude Code, and then confirmed that use in a follow-up question. Claude Code was the most commonly reported agent tool, with 86% of coding-agent users naming it; 31% reported using Codex.
Uneven adoption
Adoption varies sharply by field. Anthropic says economists reported the highest use at 39%, followed by political scientists at 25%, while public health, education, and communication were in single digits. Use also falls with seniority: just over a quarter of doctoral students and postdocs use coding agents at least weekly, but the rate is more than half that among tenured professors. The post also says researchers with typically male names were about twice as likely to use coding agents as those with female names, and researchers at top universities were 40% more likely than others to use them.
What Anthropic thinks it means
The company frames coding agents as more than writing aids. It argues they can take a research idea and dataset, write and run analysis code, interpret output, and iterate autonomously. Anthropic says that could speed up discovery, but also widen resource gaps and increase pressure on the scholarly record. It presents the survey as the first look from a larger study, including a randomized experiment on giving researchers access to Claude Code, with results to come later.
Key points
- Anthropic surveyed 1,260 quantitative social scientists in early 2026.
- 81% said they had used genAI for research, but only 20% regularly used coding agents.
- Claude Code was the most common coding agent among users, with Codex second.
- Adoption was much higher in economics and political science than in fields like public health or education.
- Anthropic says the tools could speed research, but also widen disparities and add pressure to academic publishing.
