Coding Agents in the Social Sciences

TL;DR

Anthropic's survey of 1,260 quantitative social scientists reveals that while 81% of researchers use AI chatbots, only 20% have adopted autonomous coding agents (such as Claude Code). This adoption is highly uneven, favoring early-career researchers, men, and those at high-status universities, and while these users report higher early-stage productivity, they have not yet increased journal submission rates.

The State of AI Adoption in Social Science

While a vast majority of quantitative social scientists are experimenting with generative AI, there is a significant gap between the use of general-purpose chatbots and the adoption of agentic coding tools.

General AI Use: 81% of respondents have used generative AI models to aid their research process.
Coding Agent Adoption: Only 20% of respondents regularly use AI coding assistants integrated into the command line (e.g., Claude Code, Codex, Cursor, or Google Antigravity).
Dominant Tools: Among those using coding agents, Claude Code is the most prevalent tool, reported by 86% of users, followed by Codex at 31%.

Demographic and Institutional Disparities in Adoption

Adoption of coding agents is not uniform across the academic landscape, showing steeper disparities than general AI use.

Career Stage and Discipline

Early-career researchers are the primary adopters. Just over 25% of doctoral students and postdocs use coding agents weekly, whereas the adoption rate among tenured professors is less than half that figure. Discipline-wise, economists (39%) and political scientists (25%) show the highest adoption rates, while public health (6%), education (4%), and communication (6%) show the lowest.

Gender and Institutional Status

There are statistically significant gaps (p < 0.05) in tool adoption:

Gender Gap: Researchers with typically male names have adopted coding agents at more than twice the rate of those with typically female names. This gap persists even when controlling for discipline and career stage.
Institutional Gap: Researchers at high-status and private universities are notably more likely to use coding agents.

Primary Use Cases: Code Generation over Prose Drafting

Contrary to academic concerns regarding AI-generated literature reviews and automated paper writing, the primary application of AI in social science research is technical execution.

Code Generation: This is the most common use case, reported by 97% of coding agent users and 77% of other AI users.
Prose Editing: Editing existing text is the second most common use, while drafting new prose is relatively rare; only one-third of all AI users have used these tools to draft prose.
Disciplinary Variation: Only economists and management researchers commonly use AI to draft prose.

Impact on Research Productivity

Preliminary descriptive data suggests that coding agents accelerate the early stages of the research pipeline, though they have not yet impacted the final output of journal submissions.

Early Pipeline Gains

Coding agent users report higher productivity in early-stage activities compared to researchers in the same discipline and career stage:

Project Starts: Users start approximately 0.25 more papers' worth of projects.
Working Papers: Users post approximately 0.5 more working papers.
Grant Proposals: Users report submitting more grants.

The "Last Mile" Gap

Despite gains in early-stage output, there is no evidence that coding agent users are submitting more new papers to journals or resubmitting papers more quickly. This suggests that while agents are highly effective at getting projects operational, they may be less effective at the final refinement required for journal submission.

Researcher Sentiment and Field-Level Concerns

Researchers express a dichotomy between individual productivity and the overall health of the social sciences.

Productivity Optimism: 88% of respondents believe AI makes social scientists more productive in terms of writing publishable papers.
Field-Level Skepticism: 70% of respondents are more optimistic about individual paper productivity than they are about the overall impact of AI on the social sciences. This suggests a fear that increased productivity could lead to scholarly congestion, increased competition for attention, or the exacerbation of risk-averse, incremental research.