Every major consultancy has a slide deck about reskilling. Every government has a programme. Every Fortune 500 CEO has given a speech about "investing in our people" and "preparing the workforce for the future." The narrative is comforting, bipartisan, and almost entirely unchallenged: AI will displace some jobs, but we'll retrain people into new ones. Problem solved.

Fast Company called it "the reskilling delusion" earlier this year, and the phrase landed because it names something a lot of people have been thinking but nobody in a position of authority wants to say out loud. The entire reskilling narrative might be built on assumptions that don't survive contact with evidence.

I'm not fully behind the argument. But I think it deserves more honest engagement than it's getting.

The standard pitch

The World Economic Forum's Future of Jobs Report projects that 92 million roles will be displaced by 2030, offset by 170 million new ones — a net gain of 78 million jobs. The implied message is reassuring: more jobs will be created than destroyed. We just need to help people move from the old ones to the new ones.

McKinsey estimates that up to 375 million workers globally — roughly 14% of the workforce — may need to switch occupational categories. The WEF says 39% of existing skill sets will be "transformed or become outdated" over the next five years. And 85% of employers surveyed say they plan to prioritise upskilling their workforce.

Impressive numbers. Unshakeable confidence. And the track record, if you bother to look at it, is dismal.

What the evidence actually shows

America has been running large-scale worker retraining programmes since 1962, when the Manpower Development and Training Act launched with bipartisan enthusiasm. Every subsequent programme — JTPA, WIA, WIOA — has arrived with the same promise and produced roughly the same results.

A randomised controlled trial of the JTPA programme, running from 1987 to 1992, found no statistically significant improvement in employment rates, earnings, or continuous employment for participants. Its successor, WIA, fared no better: training services "didn't have positive impacts on earnings or employment in the 30 months after participant enrolment." Participants in the Trade Adjustment Assistance programme had significantly lower employment in the first couple of years after layoffs compared to non-participants, and remained underemployed four years later.

Julian Jacobs at Brookings put it bluntly: researchers have "generally failed to show any statistically significant benefit on employment outcomes." And this is before AI enters the picture. These are results from decades of trying to retrain workers displaced by ordinary technological change and globalisation.

The pattern is consistent enough to be uncomfortable. We keep launching programmes, they keep not working at scale, and we keep launching the next one anyway.

Three problems nobody wants to name

There are structural reasons why reskilling fails, and they don't go away by spending more money or building better platforms.

Start with the job supply problem. Retraining assumes the destination jobs exist in sufficient numbers and are accessible to the people being retrained. But when AI automates customer service roles, there isn't a corresponding pool of "AI supervisor" positions waiting to absorb those workers. New roles that emerge tend to require fundamentally different capabilities: not adjacent skills you can pick up in a twelve-week course, but years of accumulated expertise in systems thinking, creative problem-solving and technical architecture.

Then there's the speed problem. The WEF's own data shows that 39% of skill sets will be outdated within five years. Corporate retraining programmes, by contrast, move at glacial speed. For the third year running, most large-scale reskilling initiatives are still stuck at the planning and activation stages, with fewer than 5% having advanced far enough to even measure success. You can't reskill a workforce in five years when the target keeps moving every eighteen months.

And then the honesty problem. A Harvard Kennedy School study by Karen Ni and colleagues found that workers transitioning to roles with high AI exposure faced a 29% earnings penalty compared to those targeting lower-exposure positions. The most rational move for a displaced worker isn't to reskill into the shiny new AI-adjacent role — it's to move sideways into something AI is less likely to touch. Which rather undermines the whole narrative of retraining people into the jobs of the future.

Where reskilling does work

I don't want to be unfair. There are cases where retraining produces real results, and they're worth understanding because they reveal what the broader programmes get wrong.

Adjacent skill transitions work. A marketing analyst learning to use AI tools for campaign optimisation isn't making a career pivot — they're extending existing expertise with a new toolset. The consultancies would call that reskilling. It's actually upskilling, and it's a fundamentally different proposition. The person already has the domain knowledge, the professional context, and the judgement that comes from years in the field. The AI tool accelerates what they can do. It doesn't replace what they know.

Small, targeted programmes with strong mentorship work. When Walmart invested $4 billion over four years to help frontline staff transition to customer-service-oriented roles, the specificity mattered. These weren't vague "learn to code" initiatives. They were moves between roles that shared enough common ground that the transition was plausible.

What doesn't work is the assembly-line-to-data-scientist pipeline that keeps showing up in policy documents. The gap between those two roles isn't a skills gap — it's a chasm. No amount of coursework bridges the difference between routine manual work and the kind of abstract reasoning, technical fluency, and self-directed problem-solving that the "jobs of the future" demand. Pretending otherwise is a kindness that helps nobody.

The comfortable lie

The reskilling narrative persists because it's useful to everyone involved. Governments can point to programme budgets. Corporations can point to training spend. Consultancies can sell the strategy. And individuals can enrol in courses and feel like they're doing something.

Nobody has to confront the possibility that some displaced workers won't find equivalent employment. Nobody has to talk about what happens when retraining isn't enough — when the honest answer is that a significant number of people will need support that looks less like a career transition, more like a social safety net.

I wrote recently about the graduation cliff — how the disappearance of entry-level roles is quietly destroying the pipeline that creates future leaders. The reskilling myth operates at the other end of the same problem. We're losing the bottom rungs of the ladder and telling the people standing on them that they can learn to fly.

The US spends roughly 0.1% of GDP on workforce development — second to last among OECD countries. A country serious about retraining doesn't spend a fifth of what its peers do. That's a country serious about the appearance of retraining.

The honest conversation

None of this means reskilling is worthless. Upskilling existing workers within their domains is valuable and should continue. Giving a project manager the tools to work with AI agents, helping a financial analyst use machine learning for pattern recognition: that works.

But wholesale career retraining, the kind the consultancy slide decks promise where displaced factory workers become cloud architects, has never worked at scale, and there's no reason to believe AI changes that. The historical evidence runs against it. The timelines are impossible. And the people selling the narrative have a financial interest in perpetuating it.

The honest conversation about AI and employment isn't about retraining. It's about what happens when retraining isn't enough. It's about social infrastructure, income support, community investment, and the uncomfortable admission that technological transitions produce winners and losers — and that the losers aren't losers because they didn't try hard enough.

Andrew Yang's proposal to shift taxes from labour to AI automation deserves attention, not because it solves the problem but because it at least acknowledges the right question. If AI is displacing workers faster than any retraining programme can absorb them, the policy response needs to go beyond training budgets. Anthropic's CEO Dario Amodei has gone further, proposing a "token tax" — 3% of revenue every time a language model generates income, redirected to government for redistribution. "That's not in my economic interest," he said, "but I think that would be a reasonable solution to the problem." When the people building the technology are volunteering to be taxed on it, that tells you something about how seriously they take the displacement risk.

The reskilling myth isn't that training is bad. It's that training is sufficient. It isn't. And the longer we pretend otherwise, the longer we delay the conversation that actually matters.

Nobody wants to have it yet. But the evidence keeps piling up that we're running out of time to start.

Quick answers

Does reskilling actually work for AI-displaced workers?

The historical evidence is poor. US federal retraining programmes running since the 1960s have consistently failed to show statistically significant improvements in employment or earnings. Where reskilling does work, it's almost always upskilling within an existing domain — not the wholesale career pivots that policy documents promise.

Why do corporate reskilling programmes fail?

Three reasons: the destination jobs don't exist in sufficient numbers, the programmes move too slowly (fewer than 5% of large-scale initiatives have advanced far enough to measure results), and the roles people are being retrained into may themselves be automated within years. The target keeps moving faster than any training programme can follow.

What's the difference between reskilling and upskilling?

Upskilling extends existing expertise with new tools — a financial analyst learning to use machine learning, for instance. Reskilling means a wholesale career change to a fundamentally different role. The first has decent evidence behind it. The second, at scale, has almost none.

What should governments do instead of just funding retraining?

Retraining budgets aren't worthless, but they aren't sufficient either. The policy conversation needs to expand to include social infrastructure, income support, and mechanisms like Andrew Yang's proposed labour-to-AI tax shift or Anthropic CEO Dario Amodei's "token tax" proposal. When AI displaces workers faster than programmes can absorb them, the response has to go beyond training.