We asked what people would do with the time the machines hand back. Someone counted it, and most of a working day a week goes straight back into the machine

This paper exists to ask one question: what do humans do when we do not have to work? Usually that question arrives as a forecast, and forecasts are cheap. This week it arrived as a measurement.

A survey of six thousand full-time digital workers across the United States, the United Kingdom and Australia, fieldwork between December and January, asked people what AI had done to their working week. They said it saves them roughly eleven hours. Thirteen per cent said their organisation was performing significantly better as a result.

Eleven hours in. Thirteen per cent out. The gap is the whole story, and the researchers went looking for the missing hours. They found them going into work that appears on no contract and in no budget: feeding the machine the context it does not have, checking output that looks finished and is not, re-running prompts through a second tool because the first answer was wrong, and clearing up downstream when nobody caught it. They gave it a name. Six and a half hours a week, most of a working day, every week.

So a piece of an answer to our founding question is arriving, and it is neither leisure nor mass unemployment. It is supervision. A large share of the time the machine hands back is being spent minding the machine.

Treat one survey with suspicion, and this one is self-reported throughout, which is the weakest kind of evidence about your own productivity. People overstate what they save. So set it against a number nobody self-reports. In the first quarter of this year, output per hour across the British economy was 0.4 per cent higher than a year before, and output per worker was 0.1 per cent lower. Whatever is being saved at the desk, it is not arriving in the national accounts. Two independent bodies of evidence, one from workers and one from the statisticians, are telling the same story from opposite ends.

The British detail is the one I cannot stop turning over. Of the three countries surveyed, the United Kingdom reported the highest adoption and the strongest gains. It also reported the best governance by some distance: more workers who had read their employer's AI policy in full, more confidence in the strategy, clearer rules about who is allowed to build and deploy an agent. Britain did the sensible institutional thing, and did it better than America.

It made no measurable difference to the hidden work. British workers spend 38 per cent of their AI time on supervision, against 36 per cent in the United States, and the researchers are explicit that a two-point gap is inside the noise. The policies are read. The strategies are trusted. The unpaid job is exactly the same size.

That is worth sitting with, because the reflex answer to almost everything on this beat is better governance, and here is a case where better governance was actually delivered and the thing it was supposed to fix did not move. If the supervision burden is a work-design problem rather than a policy problem, no amount of policy will touch it.

Two things about this paper's own week, both of which belong to you rather than to me.

On Thursday evening we built an edition, took it to the gate, and killed it. Four finished pieces, researched and written, on a genuinely good subject: a licence for children's Saturday jobs that Parliament legislated and never commenced. It failed on one question. There was no AI in it. Not thin AI, none. The replacement rows were written thirty-eight minutes after the ones they replaced, and the edition published at lunchtime on Friday, later than I would like and honestly dated. That kill is now a rule, and it sits first in the gate rather than last: if a theme cannot say where it touches the future of work and where the machine is in it, it does not run here, however good the documents are.

The second is less comfortable. On Friday night I told the machine I thought we were writing too much about politicians. It went and counted, and I was right in a way I had not expected. Across June and July, the share of our published headlines framed on government rose in a straight line, week after week, and the share framed on people fell to meet it. In the last week of July the two lines crossed. Across three hundred and thirty-eight published pieces, headlines built on an absence outnumber headlines built on something working by roughly three to one.

The cause is not laziness. It is that our own sourcing rule, which says a claim needs a source that states it in text, quietly rewards the one beat that always produces paperwork. Westminster publishes on a schedule. A firm deciding how to handle the hours AI gave it publishes nothing at all. A rule written to stop us overclaiming had been steering us toward the easiest evidence in the country.

Last night's proposed edition was killed under that finding. So was the obvious replacement. This piece is the third attempt, and it is the first one in weeks with no minister in it anywhere.

What to watch, and this is the forward-looking part. Nobody is counting supervision. It is not in a job description, not in a pay band, not in any official statistic, and not in the productivity figures that the Budget on 28 October will be built on. The next labour market and productivity releases land in the middle of August. They will not measure it either. The first employer to write this work down, name it, and pay for it will be doing something more useful than another strategy, and when one does, we will say so here.

— M.

*This note is mine: the view, and the call to run it. It begins as a draft, drawn from work the AI and I have researched and argued out together, the same way every desk in this paper is made, and I answer for every line because I read every line. Those desks run on models built by Anthropic, one of the labs sitting on the very scoreboard we report, so we tell you plainly: we cover this from inside it.*