Anthropic study finds men use AI coding agents more than twice as often as women in social science research
Another study about gender gaps in tech, and this time the culprit is coding agents. Anthropic's researchers found something that should surprise no one but still manages to sting: men use their AI coding tools at more than twice the rate of women, even when you control for career level and discipline. Economists are at 39 percent adoption. Education researchers? Four percent. The gap here dwarfs what we see in general AI usage, which tells us something important about who these tools are actual
Analysis
Another study about gender gaps in tech, and this time the culprit is coding agents. Anthropic's researchers found something that should surprise no one but still manages to sting: men use their AI coding tools at more than twice the rate of women, even when you control for career level and discipline. Economists are at 39 percent adoption. Education researchers? Four percent. The gap here dwarfs what we see in general AI usage, which tells us something important about who these tools are actually built for and who feels invited to use them.
Let's get the obvious out of the way first: this is not really a surprise. The tech industry has spent decades building tools that primarily serve the already-privileged, and coding agents are no exception. They ship with assumptions baked in—assumptions about what you already know, what problems you're trying to solve, and what kind of work matters enough to automate. If you're an economist running regressions, a coding agent is your dream. If you're an education researcher conducting interviews and analyzing qualitative data, the tool is basically looking at you blankly, wondering why you haven't tried asking it to write Python yet.
But here's where I get frustrated with the framing of these studies. The data is real, the disparity is real, and yes, we should care. But I'm tired of research that stops at documenting the gap without interrogating the machinery that produces it. Why are education researchers at four percent? Is it because they don't need coding agents, because they don't know they exist, because they tried one and found it useless for their actual work, or because the lab culture around them doesn't encourage it? These are four very different problems, and they demand four very different solutions. Throwing "more outreach" at a tool-fit problem doesn't help anyone.
The economist number is telling for another reason. Economics has spent the last two decades aggressively colonizing other social sciences with its methods. It fetishizes causal inference,RCTs, and clean empirical designs. The tools follow the culture. If your field values the kind of work that coding agents optimize for, you'll use coding agents. If your field values other forms of knowledge production, you won't. Blaming individual researchers for not adopting a tool that wasn't designed for their epistemology is like blaming someone for not wearing a hard hat to a poetry reading.
I also want to be blunt about something Anthropic probably doesn't want to hear: your own product marketing might be part of this problem. The demos for Claude, like every other coding agent, feature people building apps, automating data pipelines, and writing Python scripts. The implicit message is clear: this is a tool for builders, for coders, for people who already speak the language of software. If you're a researcher whose relationship with code is "I copied this script from a colleague three years ago and it runs when I need it to," you're not seeing yourself in those demos. The company that benefits from broad adoption has a direct interest in making the tool feel accessible to everyone, not just the people who already sound like the marketing copy.
There's another uncomfortable dimension here. The study looked at "typically male names" and "typically female names," which is a reasonable proxy but also a deeply imperfect one. It can't tell us whether the male-named economists are men, whether the female-named education researchers are women, or whether nonbinary researchers exist in this data at all. It tells us something about the names on papers, which is a useful signal, but we should be cautious about extrapolating too much about individual behavior from aggregate name-based patterns.
What actually interests me is the gap between coding agents and general AI use. Women in social science are clearly using AI—just not for coding. They're using it for literature reviews, writing assistance, maybe even qualitative analysis. This isn't a story about women rejecting AI. It's a story about a specific category of AI tool that appeals to a specific kind of researcher, and that category happens to align with the demographics of the most privileged members of academia. The tool isn't neutral. The culture isn't neutral. And pretending that better onboarding will fix this is naive.
If Anthropic wants to turn this research into something actionable, they should stop studying adoption patterns and start building tools for the researchers who aren't using them. What would a coding agent look like if it were designed for someone running focus groups with high school teachers? What if it knew what an education researcher actually does all day? Until someone answers those questions, we're just counting gaps and writing papers about them, which is, ironically, exactly the kind of work that doesn't attract coding agent users.
Disclaimer: The above content is generated by AI and is for reference only.