London wants an early-warning system for AI taking jobs. It's trying to build a smoke alarm for a fire with no flame.

Of the three things the Mayor accepted from his taskforce, the one that should interest a systems desk most is the least glamorous: a London AI Early Action System, a warning system meant to catch AI-driven change in the job market before it turns into a crisis. Its own description is that it would fold labour-market data, employer insight and local evidence into a single signal, so the city can act while there is still time. The instinct is right. The thing itself is one of the harder sensors anyone has tried to build, and it is worth seeing why.

A warning system is only as good as the signal it can catch, and a good signal is three things: fast, in one place, and easy to tell apart from everything else. A flood gauge works because water rises quickly, at a spot you chose, and a rising river looks like nothing else. Trip the alarm and you know what it means and where to go.

AI displacing work is the opposite on all three counts. It is slow: a job doesn't vanish, its tasks get absorbed one at a time, until a team of six quietly does what eight used to. It is spread out: it happens task by task across thousands of firms, not at one factory gate you can film. And it is tangled up with everything else, which is the hard part. In the taskforce's own evidence, firms described a hidden contraction in their hiring that they could not cleanly blame on AI rather than a nervous economy. If the people inside the building can't separate the two, a dashboard downstream will struggle to.

Then there is the lag. The instruments Britain already has, the monthly ONS labour and vacancy figures, are careful, which is another way of saying they tell you what happened weeks or months ago. They count jobs advertised and jobs lost. They do not count the task a piece of software absorbed inside a job that still exists. By the time a warning system built on official numbers flashes red, the change it is warning about happened several quarters back. That is a late-warning system wearing an early one's name.

None of which makes the idea wrong. Flying blind is worse, and on this, Britain largely is. But to be early in more than name, the sensor would have to do the uncomfortable thing: stop waiting for national statistics and measure the workplace directly, in near real time, task by task, what people do all day and what quietly stops being done by them. That can be built. It is also the moment a jobs sensor turns into a surveillance system, and the question shifts from can we make it to who holds the data, and who agreed to be watched.

So the most reasonable-sounding of the three proposals is quietly the most demanding. An early-warning system is a promise that you will see the thing coming. On this signal, seeing it coming means either trusting official numbers that arrive too late to be early, or building something close enough to the work to be uncomfortable. London has chosen to try. Whether it ends up with a smoke alarm or a very slow thermometer is a design choice it has not made yet.