The promise is as old as automation itself. When a machine takes a job, the worker will simply be trained for a better one, and the economy will move on richer than before. It is a comforting story, and for much of the industrial era it broadly held. The question now consuming Washington and Silicon Valley alike is whether it can survive contact with artificial intelligence, a technology that threatens to displace people faster than the country has ever managed to retrain them.

The displacement is no longer hypothetical. More than 140,000 technology workers lost their jobs in just the first five months of this year, a pace running about a third higher than the year before. Companies have now openly blamed artificial intelligence for more than 50,000 American layoffs over the past year and a half, and those are only the cuts they are willing to name. For the first time the anxiety is concentrated not in factories but in the offices, cubicles and coding teams that were supposed to be safe.

Money arrives

The response has been swift, at least in dollars. In June a new nonprofit called RAISE US was launched by Gina Raimondo, the former commerce secretary, and Eric Holcomb, the former governor of Indiana, with an explicit mission to prepare American workers for the AI economy. It has already gathered more than half a billion dollars toward a billion dollar goal, and the donors are telling. Amazon, Anthropic, Microsoft and OpenAI, the very companies building the technology doing the displacing, are helping to fund the cleanup.

There is something fitting, and something faintly uneasy, about that arrangement. The firms profiting most from automation are underwriting the effort to soften its blow, which is either corporate responsibility or a form of insurance against a public backlash, depending on how cynical one wishes to be. Either way the cash is real, and it lands on top of a federal system that has been trying to do this work, with mixed results, for a very long time.

A sobering record

That system is the problem the money cannot simply buy its way past. The country's flagship effort, the Workforce Innovation and Opportunity Act passed in 2014, funds counseling, skills assessments and training vouchers for people who have lost their jobs. On the surface its numbers look respectable, with around 70 percent of core participants employed a year after leaving. Look closer and the shine dulls, because those figures are not measured against a comparison group. Nobody can say how many of those workers would have found a job anyway, program or no program.

The older Trade Adjustment Assistance scheme, built for workers displaced by globalisation, tells a similar and sobering story. Most of its participants did eventually find work, roughly three quarters of them, but a great many landed in jobs that paid far less than the ones they had lost. Studies of workers displaced by robots found the same pattern, with many drifting down into lower paid service roles rather than up into something better. Retraining, in other words, has often functioned less as a ladder than as a cushion.

The mobility trap

This is the awkward truth beneath the optimism. American retraining is reasonably good at one thing, getting people back into some job reasonably quickly, and quite bad at another, moving them into higher paying, more durable work that the next wave of technology will not immediately threaten again. A program that returns a laid off software tester to a lower wage support role has technically succeeded and substantively failed, and the distinction is precisely the one that matters to the household paying the bills.

Artificial intelligence sharpens the trap. The whole point of retraining is to move people toward work that machines cannot easily do, yet the frontier of what machines can do is advancing so quickly that today's safe harbour may be tomorrow's target. Teaching a displaced clerk to write code made sense a few years ago. It makes far less sense now that code is among the things these systems do best. Aiming at a moving target is hard enough without the target accelerating.

What actually works

There is a bright spot in the evidence, and it points in a clear direction. The programs with the best record are the ones run with employers rather than around them, above all apprenticeships that train people for a specific job that actually exists and pays. When a company helps design the training and commits to hiring at the end, the abstract problem of matching skills to work largely solves itself. The lesson is that retraining succeeds when it is tied to real demand and fails when it is a hopeful bet on demand that may never materialise.

That is the test facing RAISE US and every effort like it. Pouring money into generic courses and vouchers will produce plenty of activity and, if history is any guide, disappointing results. Building tight links between trainers and the employers who are actually hiring is harder, slower and less telegenic, but it is the only version that has ever reliably worked. America has the resources and now the attention to retrain its workers for the age of AI. Whether it has the patience to do it in the way that works, rather than the way that looks busy, is the question that will decide how many people the technology leaves behind.