The career ladder still stands, but its bottom rung has been quietly removed — and most of the public conversation about AI and jobs is happening at the wrong altitude to notice.

Ezra Klein's New York Times piece this weekend, "Why the A.I. Job Apocalypse (Probably) Won't Happen," is one of the more grown-up things written on AI and the labour market in recent months, and the broad argument is right. We're not heading for mass unemployment. The economists Klein leans on — Alex Imas at Chicago, Eldar Maksymov at Arizona State — make the case calmly: something is always scarce, and as automation makes one thing cheap, demand shifts toward what can't be automated. The Jevons Paradox doesn't just apply to Victorian coal. It applies to spreadsheets, to computers, and almost certainly to AI.

So far, so reassuring. The trouble is that Klein's most important point is the one he leaves until the last few paragraphs, and it deserves the whole article.

His real warning is this: a world where AI displaces eight million workers may be harder to handle than a world where it displaces eighty million. A total shock would force us to restructure the economy, as Covid did. A partial shock — the kind we actually face — gets met with a few months of unemployment insurance, some retraining schemes that don't work, and then a quiet decision to ignore the people affected. We've seen this film before with the China shock and post-industrial Britain — and the fiscal room to handle it has only shrunk in the years since.

This is the bit we should be talking about. From inside a working life that's already been restructured around AI, it looks even more pressing than Klein lets on.

Where the optimists are right

Imas's relational-economy point is genuinely strong. As automation makes goods and information cheaper, the human element becomes scarcer and more valuable. People want clothing with a story, doctors who make house calls, tutors who know their children. The more our digital lives are mediated by machines, the more we pay for the moments that aren't. There are more baristas now than before Nespresso, and more coffee shops, not fewer.

Maksymov adds a useful piece to that. VisiCalc didn't put accountants out of work; it quadrupled them, because cheap spreadsheets unleashed demand for financial work that had been priced out of the market. Klein extends this to his own working life, and so can I. Every AI-augmented person I know is working harder than ever, not less. The tools didn't take the work away — they expanded what was possible to attempt. I'm shipping more, the questions I'm able to ask have multiplied, and the answers are more ambitious. The graduate I would have hired five years ago to do research summaries doesn't really exist as a role any more, but I'm not less busy because of it.

That's the optimistic case, and it holds for most of the economy most of the time.

What the optimists understate

Where the argument gets thin is at the entry-level rung. Klein's economists are mostly right about how much work there is. They're less right, I think, about how that work reshapes itself inside an organisation. AI is displacing the bottom of the job, rather than the job itself. The graduate research role. The junior analyst clipping data. The trainee paralegal reviewing contracts. The first-year associate writing summaries. These were the rungs people climbed up. They were also where institutional knowledge accumulated — where someone learned which questions mattered, which sources got trusted, what the senior people actually cared about versus what they said they cared about.

Take those rungs out and the ladder still stands, but nobody can get on it. I've argued elsewhere that AI is already replacing jobs, software development first — what's new here is that it's now eating the bottom rung rather than the whole role.

This is the partial-displacement story Klein flags but doesn't follow through. One in five companies in a recent Resume.org survey have already stopped hiring entry-level workers because of AI. Korn Ferry's 2026 report found more than half of leaders plan to add autonomous agents to their teams this year, with 37% planning to replace entry-level roles outright. The graduate unemployment rate in the US is 5.7%, and Bill McDermott of ServiceNow argued in March it could reach the mid-thirties within two years. McDermott may well be wrong on the number — that prediction sits at the very alarmist end and I'm not convinced by it. But the trend underneath isn't in serious dispute.

The displacement here is gradual rather than sudden — role by role, hire by skipped hire. It doesn't show up in the unemployment rate. It shows up in the people who aren't there in 2036: the senior partners and team leads who never made it through the rungs that no longer exist. As I've argued before, you can't backfill institutional knowledge or fast-track judgement, and you can't hire your way out of a talent gap you spent five years creating.

Why this lands harder in Britain

Klein writes from a US perspective, where unemployment is 4.3%, the labour market is broadly tight, and software engineers are still in demand. The British picture is different.

Our productivity problem predates AI by twenty years and shows no sign of resolving itself. The management capability gap is the bigger structural worry: 82% of UK managers have no formal training, which means most of the people deciding which junior roles to cut and which AI tools to deploy are doing so without the foundational skills to make those decisions well. AI is more likely to widen that gap than narrow it — the managers who already know what good looks like will get sharper with these tools, and the ones who don't will simply have a faster way of producing average work. And the fiscal position gives the government almost no room to absorb the cost when partial displacement starts to bite.

The relational sector Imas points to as the natural home of human labour is exactly the part of our economy — care, hospitality, tutoring, personal services — where wages have been most suppressed for the longest. Klein's observation that we get crueler when displacement is more limited is describing something Britain has been particularly good at for forty years. The China shock hit specific towns. Deindustrialisation hit specific regions. We responded with regeneration schemes that mostly didn't work, and then mostly forgot. Walk around the affected places now and the wage gap is still there, the political consequences still unfolding.

If AI displacement follows the same pattern — concentrated in particular roles, particular sectors, particular age cohorts — we should expect the same response. Some retraining money, some warm words, then a slow turning away while the affected people work out for themselves what to do next.

What working with these tools daily reveals

Two things about using these tools every day don't show up in the academic argument.

The first is that the productivity gain is uneven across roles, and it compounds the wrong way. AI makes a senior person more productive in a way that strengthens them. It makes a junior person more productive in a way that often substitutes for the learning they would otherwise be doing. A senior engineer using Claude Code gets a force-multiplier on judgement they already have — they know which suggestions to accept, which to challenge, where the tool is going to go wrong. A graduate using Claude Code accepts the plausible answer, ships it, and never has the painful experience of working out from first principles why a particular approach was wrong. The senior gets sharper. The junior gets dependent. Over time, that asymmetry compounds in the direction of fewer junior hires, because the senior person genuinely needs them less and the juniors who do exist are demonstrably less capable a year in than their predecessors were.

The other thing is harder to quantify. Imas's relational economy doesn't materialise on its own. It needs people who can do relational work, and as Klein notes near the end of his piece, those skills are getting rarer in exactly the cohort that will need them. Young people spend less time with friends than they did in 2003, less time on dating, and far more time interacting with screens than with each other. The pattern is well-documented. AI accelerates it rather than reverses it. We may end up with an economy that needs more genuinely relational human work and fewer humans equipped to do it.

That's the bit I find genuinely worrying. Not the headline number, but the composition of what's left.

Practical responses for organisations, individuals and policymakers

None of this is useful unless it goes somewhere. Here are the moves that seem to me to matter.

For organisations, before you cut another junior role, ask who replaces the senior person who leaves in three years. If the answer is "we'll hire externally", you've described the problem, not the solution. The companies handling this well treat AI-augmented junior roles as an investment in the kind of human capability AI makes more valuable, not less. They pair graduates with the tools from day one. They protect time for the messy first-principles work that AI would otherwise do for them. They make their senior people responsible for actively developing juniors rather than just reviewing their output. And they resist the temptation to use AI as a hiring freeze with extra steps, because the hiring you skip in 2026 is the management bench you don't have in 2031.

For individuals early in their career, the strategic move is to optimise for the skills AI makes scarcer rather than the skills it makes cheaper. Judgement is one. The willingness to take real responsibility for an outcome rather than defer it to a tool is another. The ability to sit in a room with another human and read what's actually happening is a third. Use AI heavily, but treat the work AI does for you as something you should be able to redo yourself if it disappeared tomorrow. Build a portfolio of work that demonstrates you can think, not just prompt. None of these show up well on a CV, and they aren't glamorous. They'll be the ones that compound.

On policy, the lesson of every previous partial-displacement event is that we wait too long, do too little, and target the wrong thing. The British answer to AI labour displacement shouldn't be retraining schemes for jobs that no longer exist. It should be active labour-market support for the relational sector — care, education, mental health, community services — where the work is genuinely human and the pay should reflect what it's actually worth. That's a fiscal choice, and it probably means raising taxes. Whether the political appetite for that exists in 2026 is, frankly, doubtful — but pretending the problem will resolve itself through aggregate demand has a worse track record than any policy this country has tried.

The bottom line

What we face here is something harder to see and harder to organise against than mass unemployment — a slow restructuring of who gets to enter the labour market in the first place, and on what terms. In the version of this we're actually living through, the thing becoming scarce is the chance to start a career at all.