The story we keep being told about AI and work is a story about the bottom. The warehouse picker, the call-centre operative, the data-entry clerk, the junior coder fresh out of university — these are the people we're told to worry about, and the remedy we're offered is always the same. Retrain them. Move them up. Teach them to do the work the machine can't yet do. It's a tidy narrative, and almost everyone in a position of authority repeats it, because it lets the disruption feel like a problem with a solution.

I think it's mostly wrong. Not the part about disruption — that's real, and I've spent a fair amount of time arguing that it's real. The part that's wrong is who we keep pointing at. The damage is landing somewhere other than where we're all looking. It's landing in the middle. And the remedies we keep proposing are either hollow or, in a lot of cases, theatre.

Start with where the harm actually falls. The early consensus held that AI would eat the bottom rung first, and there was a logic to it: the roles closest to the work AI already did well looked the most exposed. Some of that has happened. But the heavier blow is landing on the layer above — the coordinators, the schedulers, the people who translate a strategy into a set of tasks and chase them to completion. That work turns out to be exactly what AI agents are getting good at fastest. A middle manager isn't easier to replace than a junior; the difference is that so much of what a middle manager does is coordination, and coordination automates beautifully. I've written before about how AI is unbundling middle management, splitting the coordination work from the leadership work and absorbing the first. Watch what happens when an AI system stops completing tasks and starts assigning them, and you start to see the shape of it. The hollowing happens in the middle, well above the floor.

This matters because the middle is where careers are made. The middle is the rung you climb to on your way to anything senior. Take it out and you don't just lose a tranche of jobs this year — you lose the path that turns a junior into a leader a decade from now. We're removing the staircase and telling people to jump.

Which brings me to the remedy, the one everyone reaches for: reskilling. Every consultancy has the slide deck, every government has the programme, every chief executive has given the speech about investing in our people. And the evidence for it, if you actually look, is dismal. America has been running large-scale worker retraining since 1962, and the controlled trials keep returning the same answer — no statistically significant improvement in earnings or employment for the people put through it. I've gone through this at length in the reskilling myth, so I won't relitigate the whole case here. The short version is that upskilling works — giving a marketing analyst AI tools to do the job they already understand — and wholesale retraining, the assembly-line-worker-to-data-scientist pipeline that fills the policy documents, has essentially never worked at scale. The gap between those two roles is a chasm, not the kind of thing you close with a twelve-week course. “Learn to code” was the same promise a decade ago, and it aged badly. “Learn to prompt” will age worse.

The reason the reskilling narrative survives despite the evidence is that it's useful to everyone telling it. Governments get to point at a budget line. Corporations get to point at training spend. Consultancies get to sell the strategy. And the displaced worker gets to enrol in a course and feel they're doing something. Nobody has to say the harder thing, which is that some people genuinely won't find equivalent work, and that the answer to that looks less like a career transition and more like social infrastructure. The comfort of the narrative is precisely the problem. It lets everyone off the hook.

And then there's the layer of theatre on top of all this, which is the part that irritates me most. A great many of the layoffs being blamed on AI have very little to do with AI. They're the cost-cutting a company wanted to do anyway, dressed up in the language of transformation because the language is rewarding. I dug into this in the layoff theatre: a chief executive who announces redundancies because margins need fixing gets a sympathetic nod and a modest bump; the same chief executive who blames the cuts on AI gets a stock surge and an invitation to Davos. Even Sam Altman, who has more to gain than almost anyone from the everything-is-being-transformed story, has admitted there's “AI washing” going on — companies pinning ordinary layoffs on a technology because it plays well. When the person with the strongest incentive to overstate AI's impact is telling you it's being overstated, that's worth hearing.

None of this is a conspiracy. It's just how incentives work. A restructuring story that used to be told as “the internet didn't deliver” or “post-pandemic normalisation” now gets told as “AI transformation,” and the layoffs happen at roughly the same rate they always did. The fashionable explanation changes; the spreadsheet underneath doesn't. The trouble is that the theatre does real damage. It poisons the well for the genuine cases — the companies actually rebuilding around the technology get lumped in with the performers — and it makes an honest conversation about the real transition almost impossible to have. When everything is AI, nothing is, and the people who should be adapting quietly conclude it's all hype.

The people building the technology have noticed which way the political wind is blowing, and they're already trying to write the response. OpenAI has published a thirteen-page policy paper proposing robot taxes and citizen wealth funds — and when you read it closely, almost none of the proposals would cost OpenAI anything. I picked through it in Sam Altman's new deal. The robot tax falls on the adopters, not the model provider. The reskilling assumption the paper leans on is the one we already know doesn't work. It's a positioning document dressed in social-democratic clothing, and it leans on the same comfortable story about retraining that the evidence keeps refusing to support.

So what does an honest version of this look like? It starts by aiming at the right target. If the squeeze is in the middle, then the policy conversation about graduate hiring, about career ladders, about the disappearing rungs is more urgent than another round of retraining grants. It means being honest that upskilling within a domain is worth doing and wholesale career reinvention mostly isn't, and not pretending otherwise to spare ourselves a difficult conversation. It means applying a simple test to every layoff that arrives wrapped in the language of transformation: can the company name the specific capability that changed the economics, show the before-and-after numbers, and point to the AI talent it's hiring even as it cuts? Most can't. The ones that can are doing something real, and they're rare enough to notice. I'm not sure what the right ratio of genuine to theatrical actually is — nobody is, including the executives making the announcements — but the test at least sorts the two.

And for anyone living inside this rather than theorising about it, the move is the same one I'd give a friend: don't adopt the tools reflexively because you've been told to, adopt them where they genuinely make you better and guard the parts of the work that are actually yours. I wrote a whole piece on surviving an AI mandate for exactly this reason. The worker who uses AI to do more of the same becomes more replaceable; the one who uses it to do work the machine can't becomes harder to lose.

The disruption is real. I want to be clear about that, because the argument I'm making could be misread as a shrug, and it isn't one. AI is genuinely changing how a lot of work gets done, and pretending otherwise helps nobody. But the version of the story we keep being sold — that the danger is at the bottom, that retraining will catch the fallen, that every layoff with “AI” in the press release is a glimpse of the future — gets almost every part of it wrong. The harm is in the middle. The remedy is hollow. And a fair slice of the disruption is theatre. We could have a serious conversation about any of it, but we'll only get there once we stop pointing at the wrong people.