APREX · EST. 2026
The chatbot is common. The agent is not.
Anthropic has mapped how social scientists use AI in their work. The numbers reveal a clear divide: ordinary chatbots are already widely embedded in research, while code agents remain in early adoption.
The survey was conducted in February and March 2026 and covers 1,260 social scientists. 81 percent said they had tried AI chatbots in their research. Only 20 percent had adopted code agents — tools in the vein of Claude Code.
AI is already inside research, but the most productive agent use is still unevenly distributed.
Why code agents represent a different leap
A chatbot can help with text, ideas, and code drafts. A code agent can go further: write analysis code, run it, debug, iterate, and produce results that feed directly into the research itself.
Anthropic reports that the most common AI use among both code-agent users and other AI users is generating code for quantitative analysis. They also report that 97 percent of code-agent users and 77 percent of other AI users use AI for code generation — which says a great deal about how central programming has become in social science work.
Unequal access becomes a research-policy question
The most uncomfortable finding is not simply the adoption gap. Anthropic reports that researchers with typically male names are twice as likely to use code agents as those with typically female names, and that researchers at top universities are 40 percent more likely to use such tools.
This is not merely a productivity footnote. If code agents genuinely make it easier to write analyses, produce working papers, and submit grant proposals, unequal access could amplify existing inequalities in academia.
The broader research trajectory
This comes as agent research is moving fast. The Nature publication on The AI Scientist presented a system capable of proposing ideas, writing code, running experiments, drafting manuscripts, and undergoing automated review in machine-learning research. The researchers stressed clear limitations, but the point stands: the research process itself is beginning to be agentified.
The Anthropic study is not about fully automated research. It is about human social scientists. That is precisely what makes it important: before full autonomy arrives, we are already seeing unevenly distributed assistance.
Conclusion
Anthropic's study shows that AI adoption in research is not a single phenomenon. Chatbots are nearly normalised, while code agents remain an early-stage and unevenly distributed productivity tool.
For universities and research institutes, the lesson is clear: do not wait until agent use becomes an informal elite competency. Make tools, rules, and training broad enough that the productivity gains do not accrue only to those who already have the most.