Men Use AI Coding Agents at Twice the Rate of Women. Four Theories Try to Explain Why.
By Courtney Turrin · August 17, 2026 · 11 min read read
Category: For Curious People
Men use AI coding agents at twice the rate of women, a gap that widens beyond simple chatbot use. This trend may impact entry-level hiring.
In February 2026, Anthropic gave a group of quantitative social scientists free access to Claude Max and tracked what they did with it. Eighty-one percent had already tried an AI chatbot for their research. Only 20% had adopted a coding agent, a tool that can write, run and revise analysis code with minimal supervision, into regular use. Researchers with names classified as typically male adopted the agents at more than twice the rate of researchers with typically female names, and the gap did not shrink among people who had already tried some form of AI. It was slightly larger.
That gap sits one level above where most coverage of AI and gender stops looking.
Chatbot Use Reached Parity. Agent Use Did Not.
A Pew Research survey of more than 5,000 U.S. adults, conducted in February 2026, found 50% of men and 47% of women say they have used an AI chatbot, closing what had been an 11-point gap in 2024. That is the figure that gets quoted. It also describes the lowest possible bar for "using AI."
Move one rung up and the gender gap in AI adoption reopens. The same Pew survey found 27% of men use chatbots daily, against 20% of women. Men report higher confidence with the tools, 22% saying they are extremely or very confident versus 15% of women, and men are more likely to say they have heard a lot about AI in general, 55% to 41%.
Move up another rung, to building rather than chatting, and the gap stops looking like a gap and begins to resemble a chasm. A working paper synthesizing 76 sources across more than 100 countries and over 300,000 respondents puts the overall adoption rate at 47.8% of men versus 39.3% of women, a relative gap that has held around 16% since early 2025 (Fig 1). The same paper cites the Anthropic coding-agent finding directly, describing it as evidence that a second, more durable kind of gap opens once tools move from answering questions to producing work.
[datawidget key="ai-adoption-ladder"]
The Real Stakes Are Hiring, Not Pay
For most of the past three years, the case for closing the AI gender gap rested on pay: if AI makes people more productive and productivity gets rewarded, people using it less would eventually earn less. That argument has gotten harder to make. A National Bureau of Economic Research paper linking adoption surveys to administrative labor records for 25,000 workers across 7,000 Danish workplaces found no significant effect of AI chatbot use on earnings or hours across 11 exposed occupations.
Whether AI is behind the broader wave of 2026 layoffs is a separate question, and Channel RVA has already reported that case closely. That reporting found the link between AI and mass layoffs thin, but found one place the effect shows up clearly: workers in their first jobs. Stanford's Erik Brynjolfsson, Bharat Chandar and Ruyu Chen, using payroll data from ADP covering roughly one in six U.S. workers, revised their estimate in August 2026 to find employment for workers ages 22 to 25 in the most AI-exposed occupations now stands 19% below where it would be had it kept pace with less-exposed peers, a gap that has widened since they first documented it in 2025. The same researchers state directly that they find no evidence of widespread, economy-wide job displacement. The effect is concentrated at the entry-level rung specifically, which is exactly where career-defining AI fluency would need to show up first.
That is where the gender data and the hiring data intersect. McKinsey and Lean In's 2025 Women in the Workplace report, drawing on roughly 9,500 employees surveyed across more than 120 companies, found only 21% of entry-level women say their manager encourages them to use AI tools, compared with 33% of entry-level men (Fig 2). Entry-level women were also far less likely to believe AI would improve their career prospects: 37%, against 60% of employees overall. If entry-level hiring is the rung where AI exposure most directly affects who keeps a foothold in a career, it is also the rung where the encouragement gap is widest.
[datawidget key="manager-encouragement-gap"]
Four Competing Explanations for the Agent Gap
Ask why the gap exists and the explanation that circulates most widely is that women anticipate being judged more harshly for using AI. There is evidence behind this hypothesis. Researchers had engineers evaluate identical code, varying only whether the author was described as male or female and whether AI assistance was disclosed. Harvard Business Review reported that reviewers rated perceived AI users 9% less competent on average, with the discount at 13% for women and 6% for men. Male engineers who had not adopted AI themselves penalized women 26% more than men for identical AI-assisted work. Separately, Lean In's research found women were 22% less likely than men to be regular AI users at work and 32% more likely to worry that using AI would be seen as cheating.
That is a real, measured effect. Whether it is the main driver of the adoption gap is a separate question. The link between the penalty and lower use comes from surveys asking people whether they anticipate judgment, not from an experiment that varied the penalty and measured whether adoption moved. Three other explanations have at least as much evidence behind them, and each predicts something different.
The first is risk. A 2026 study in PNAS Nexus surveyed roughly 3,000 respondents in the U.S. and Canada and randomly assigned them to scenarios where the odds that a company's AI adoption would produce net job gains ranged from certain to a coin flip. At full certainty, the gap between men's and women's support for adopting AI disappeared. As certainty fell, women's support dropped faster than men's. The researchers tie this to two measured traits in their sample: women scored higher on general risk aversion and higher on exposure to AI-related job risk based on their occupations. This predicts the gap should shrink wherever outcomes are more certain and widen wherever they aren't, regardless of who is watching.
The second is self-assessment. In a well-known experiment, sociologist Shelley Correll found that exposing participants to a stated belief that men perform better at a task changed how participants rated their own competence at that task, independent of their actual measured performance, and those self-ratings then predicted career aspirations. Applied here, the prediction is that the threshold for "I am ready to build with this" sits higher for women, not because of what anyone else will think, but because of an internalized bar. This would show up even in private, unwitnessed use, which the anticipated-penalty explanation would not fully predict.
The third is conditioning, and it starts earlier than any of the others. In a study of 298 families at children's museum exhibits, researchers found parents were three times more likely to offer causal, scientific explanations to sons than to daughters, despite no difference in how often boys and girls asked questions or how often parents discussed simply operating the exhibit. The gap was specifically in explaining the mechanism, not in showing children how to use it. The same asymmetry has since been observed with physics toys at home and shared reading of science books.
That distinction, between being shown how to operate something and being taught how it works underneath, maps onto the difference between using a chatbot and building with an agent. It may also help explain one of the largest documented gender differences in psychology. A meta-analysis of 47 interest inventories covering more than 500,000 respondents found men score substantially higher on interest in working with things over people, engineering interest chief among them. That is a large effect, and it is also, by the authors' own account, sensitive to how the inventory was built: item-development choices alone account for roughly a fifth of the variance in the size of the gap across the 47 instruments studied. A vocational preference measured in adults is not obviously separable from what those adults were taught to find interesting as children.
Four explanations, then, each making a different prediction: the gap should close as certainty rises if it is about risk; persist even in private if it is about self-assessment; track early exposure to explanation over instruction if it is about conditioning; and track visibility and scrutiny specifically if it is about anticipated judgment. Anthropic's coding-agent survey is the baseline wave of a larger study that will eventually give some researchers randomized access to Claude Code, which should help separate these explanations. For now, more than one of these mechanisms is likely operating at once.
Planning First or Building First: An Unsettled Question
Set the explanations aside and look at what happens once someone sits down to build. Anyone who has watched people work in a tool like Lovable or Replit has likely noticed two different postures: one that spends most of its time in the planning stage before writing a prompt, and one that starts generating immediately and iterates through what breaks. Whether this splits cleanly along gender lines, in the way the risk-orientation research would predict, is not something this reporting could confirm. It would take a structured comparison across a real sample of builders, not an impression from watching a handful of people, and no one appears to have run that comparison yet.
What is measurable is which posture the current evidence favors. In July 2025, METR ran a randomized controlled trial with 16 experienced open-source developers completing 246 real tasks in codebases they had maintained for years. Developers using AI took 19% longer than those working without it. Before the study, they had predicted AI would make them 24% faster. Afterward, having experienced the slowdown, they still estimated they had been 20% faster than they actually were.
METR tried to repeat the study in 2026 with newer tools and could not. In a February 2026 update, the researchers reported that too many developers refused to be randomly assigned to work without AI at all, and that how developers worked with the tools had changed enough to make simple time-on-task comparisons unreliable. The finding that survived both rounds was not a verdict on whether AI makes developers faster. It was that developers misjudge their own AI-assisted output, overestimating their speed in the 2025 trial and, by the researchers' own account, likely underestimating it in 2026, and that a snapshot of an early tool used by early adopters is not a durable measurement of what the tool can do.
That instability is itself an argument for treating upfront planning as a skill rather than a hedge. A 2026 paper on AI-assisted skill formation describes what researchers call a jagged capability frontier: AI improves performance on tasks inside its current competence and degrades it on tasks just past that boundary, and workers are generally poor at perceiving exactly where the boundary sits. Someone who plans thoroughly before prompting, arriving with a clear idea of what the tool should produce, is doing something close to testing that boundary before committing to output. Someone who ships first and fixes what breaks is betting the errors will be cheap and visible. Both strategies have documented failure modes. The first can look slow next to someone building in public. The second is the exact pattern METR's developers showed when they believed they were 20% faster while measurably running behind.
The Case That Caution Is the Correct Call
None of the above should be read as more settled than it is. Nearly every figure in this piece rests on self-reported survey data. Women are more likely than men to answer "I don't know" on surveys about ambiguous or technical topics, which could inflate an apparent gender gap in uncertainty about AI. The authors of the PNAS Nexus study addressed this directly. They argue the finding reflects real uncertainty rather than a response-style artifact, because the same survey found men scored higher on objective, factual questions about AI, and lower objective knowledge is independently associated with higher perceived risk.
There is also a more basic possibility: caution is simply the correct call given what is currently known. METR's own history is the clearest illustration available. The industry consensus in 2025 was that AI made developers faster. The first time anyone tested that claim under controlled conditions, it turned out to be backward, and the developers using the tools did not know it. If expert judgment about a fast-moving technology can be wrong at the level of professional software engineers running a randomized trial on their own work, a posture of withheld confidence toward these tools is at least as plausible a reading of the evidence as an error.
What the frontier layer of AI use rewards next is still open. Stack Overflow's 2026 developer survey found AI tool adoption at a record 84%, alongside a record low in trust of AI-generated output, with only 3% of developers reporting they highly trust it. Microsoft's 2026 Work Trend Index found the human skill AI users ranked most important going forward was quality control of AI output, ahead of critical thinking and everything else on the list.
If verification and judgment turn out to be the scarcer, more valuable skill as these tools mature, the shape of who is currently ahead and who is currently behind may look different in three years than it does from inside 2026.