We are promised that retraining will carry people through the AI shift. The best evidence says it barely moves the needle.

Start with the promise, because it is a good one. Reskilling, done well, pays. A widely cited McKinsey study put the productivity gain from effective reskilling at six to twelve per cent, and found it made economic sense for UK employers in roughly three-quarters of cases. That is the number that ends up in strategies and speeches. It is also the kind of number the advisers to the incoming Prime Minister leaned on when they briefed a newspaper this month about a reskilling guarantee for workers exposed to AI.

Now set the promise against the measurements, which are a great deal more sober.

A meta-analysis of adult retraining programmes across many countries found that, on average, taking part raised a person's chance of finding a job by about two and a half percentage points, and nudged wages up by a fraction. Real, but small. A UK what-works review of support for displaced workers found that training often has limited short-term effect, and can even lengthen the time someone spends out of work, because retraining takes them off the market while they do it. It helps most, that review found, when it is tightly targeted at people who have been out of work a long time or are deliberately changing career, not sprayed across a workforce as a reassurance.

The most careful recent survey of the evidence, from the Brookings Institution, is blunter still. It warns against assuming public retraining will protect people's current jobs or reliably find them new ones, and calls the evidence on retraining efficacy inconclusive at best. Then it makes the point that matters most for us. Even the training schemes that do work may only change which workers get access to the good jobs that already exist. They do not create more good work. They reshuffle the queue; they do not lengthen the counter.

Hold that next to the thing this paper keeps saying. Every retraining promise, the good ones included, assumes a destination: a stable, better job on the far side worth crossing to. That assumption held, more or less, when the machines were taking particular tasks and leaving whole occupations standing. It is exactly the assumption AI puts in doubt, because the jobs you would retrain people into, the analysis, the drafting, the first-line judgment, sit on the same list as the ones they are leaving. When the destination recedes as fast as you send people toward it, a couple of percentage points on their odds of arriving is not a plan. It is a rounding error against the size of the problem.

None of this says stop retraining people. Targeted, patient, honest retraining helps real people, and the evidence for that is solid. It says something narrower and harder. Retraining is a bridge, and a bridge needs two banks. The measured reality is that we are building furiously from one of them, and nobody in authority is checking whether the other one is still there.