The home computer story from the 1980s has a specific shape. A new technology arrives. Access to unstructured time with it is unevenly distributed. A few years later, differences in fluency are read as differences in aptitude, and institutions calibrate to the fluent group.
Generative AI tools are new enough that the same process, if it is occurring, is currently observable rather than historical. That makes it worth watching carefully and worth being specific about what would constitute evidence.
What is actually known so far
Considerably less than the volume of commentary implies. This is a fast-moving area, survey data ages quickly, and much of what circulates is vendor material rather than research.
A few things have been reported with reasonable consistency across multiple surveys in different countries.
Self-reported use of generative AI tools has been found to be somewhat higher among men and boys than women and girls in several large surveys, with gaps varying by country and by the specific tool.
Self-rated confidence in using these tools shows a larger gap than reported use — the familiar pattern in which the confidence difference exceeds the behavioural difference.
Stated concern about accuracy, ethics and appropriate use has been reported as higher among women in several surveys.
What is not established is whether any of this translates into differences in capability, or whether it persists as tools become more ubiquitous.
Why the confidence-exceeds-use pattern is the one to watch
Because that is exactly the pattern that preceded the computing reversal.
The problem in the 1980s was not that girls could not use computers. It was that introductory courses assumed a baseline of prior tinkering, that students without it interpreted the resulting difficulty as evidence about themselves, and that the field's culture came to treat early self-taught fluency as the marker of belonging.
The equivalent risk now is a workplace or classroom culture in which fluency with these tools becomes an informal credential, acquired mostly through unstructured play, and in which people who acquired it are treated as naturally suited to technical work.
Nothing about that inference would be valid. It was not valid the first time either.
The specific asymmetry to watch for
There is a plausible mechanism worth naming, though it is currently hypothesis rather than finding.
Effective use of these tools requires a willingness to make confident but unverified assertions, observe that they are wrong, and iterate. The interaction rewards fast, low-stakes attempts and punishes careful deliberation before acting.
If girls have on average been more strongly trained toward accuracy before assertion — which is what the calibration and apology literatures suggest — then this interaction style is one they have had less practice at, and the tool will feel less natural for reasons that have nothing to do with technical ability.
This is testable and, as far as this publication is aware, has not been properly tested. It should be.
The accuracy concern is not a deficit
Surveys reporting higher concern about AI accuracy among women tend to frame it as reticence to be overcome.
Given what is now well documented about these systems generating confident falsehoods, higher concern about accuracy looks less like a barrier and more like appropriate calibration. The same pattern appears in the confidence literature: the more accurate assessment gets classified as the deficient one because the less accurate one is more assertive.
A person who checks output before relying on it is not behind. She is doing the thing the documentation tells everyone to do.
What schools should be doing differently
The lesson from the computing history is precise: do not let unstructured prior exposure become the entry requirement.
That implies teaching tool use explicitly rather than assuming it. It implies not treating students who arrived already fluent as the benchmark for the class. And it implies making the iterative, error-tolerant interaction style an explicit taught skill rather than a disposition students are assumed to have or lack.
The universities that split introductory computing by prior experience solved a version of this problem and published results showing it worked. That precedent is directly applicable.
The reason to be cautiously optimistic
Two differences from the 1980s are worth noting.
These tools are language-based, and the entry barrier is much lower than programming ever was. There is no syntax to learn and no environment to configure. The gap between novice and expert use is real but smaller.
And the pattern is documented this time. In 1985 nobody was watching the demographics of home computer access with any concern. There is now a reasonably well-developed literature on how this goes wrong, which is a genuine advantage.
Whether it is used is a separate question. The failure mode is not that anyone decides to exclude anyone. It is that a group with a head start becomes the default assumption, and nobody notices until the enrolment figures turn.