Most people learning AI are teaching themselves. That sounds like good news. Look closer.

When the Mayor's taskforce asked Londoners how they had learned to use AI, the answer was not "on a course." It was "on my own." In its engagement work, 83% said they had used AI in the last six months, nearly half most days, and around three quarters said they had picked it up by teaching themselves and experimenting. Formal education barely featured. People are not waiting to be trained. They are poking at the thing until it does something useful.

That reads like the easiest possible backdrop for a skills programme: a population already leaning in. It isn't, and the reason is who answers a survey like this in the first place. The people who fill in a mayor's AI questionnaire are, almost by definition, the ones already curious and confident about it. The exercise measures the enthusiasts. The quieter truth sits with everyone who didn't reply.

Because whether you teach yourself a new tool has less to do with aptitude than with something plainer: slack. Whether your day has ten spare minutes to fail in. Whether your job rewards you for trying something new or marks you down for looking unsure. Whether you carry the quiet sense that this sort of thing is "for people like me." The self-teachers aren't cleverer. They mostly had more room.

Now set that against who the taskforce says is most exposed: 600,000 Londoners in clerical and administrative work, mostly women, paid below the average. That is often work with the least slack in the day and the closest watch on the small mistake, the two conditions under which nobody experiments. The people with the most reason to learn AI are the ones whose working lives make teaching themselves hardest.

So "most people teach themselves" is not the reassurance it sounds like. It is a warning about the ones who don't, and why. A reskilling offer built for a self-teaching population risks reaching the people who would have got there anyway, and missing the ones the map calls exposed.

The deeper thing the behaviour shows is how people absorb a technology that lands in the middle of a working life: not in a classroom but in the gaps of the day, unevenly, gated by confidence as much as capacity. Britain measures how tight those gaps already are in the ONS Time Use Survey, and they are tightest for women, who still carry more of the unpaid work that fills them. A programme that wants to reach the exposed may have to fund the un-obvious thing: not more training, but the time and the permission to be bad at something for a while. That is harder to put in a press release than a course. It is also more honest.