Raising the floor is not building a ladder

One of these is my argument. The other is my evidence. They disagree.

I’ve been making a version of this argument for a while: as AI absorbs routine production work, the highest-value editorial work moves upstream into evidence, judgement and narrative.

There’s an uncomfortable corollary I’ve written about but not resolved. The routine production work AI absorbs first is also how people used to learn the job.

Nobody designed that apprenticeship. It happened as a by-product. You wrote the fifth version of a page nobody read and noticed why the first four didn’t land. You transcribed the interview and found the useful answer arrived halfway through an unrelated question. You produced something competent, watched it do nothing, and were asked to explain why.

If that work goes, the ladder goes with it.

That’s the argument. Then I looked at my own evidence.

What the evidence actually said

I ran an AI-assisted editorial workflow end to end on a real project, and recorded the outcome at the time rather than reconstructing it later.

The strongest practical result was this: the workflow let a less experienced writer produce a credible first draft, ready for editorial review. It did not make an experienced writer’s best work better.

I wrote it down as the workflow raises the floor rather than the ceiling, and marked it observed once.

That is not what my argument predicts.

If AI mainly compounds the advantage of people who already have judgement, the experienced writer should have gained most. They didn’t. The person who gained most was the one with the least accumulated experience — precisely the person the ladder argument says is being left behind.

One observation, one project, one writer. I would not build anything on it. But it is the only direct evidence I have, and it disagrees with me.

Both can be true, and that’s worse

The obvious move is to pick one. I don’t think you can, because the two claims are about different things.

Raising the floor is a claim about output. With good scaffolding and accumulated context, someone early in their career can produce work that reaches a professional standard much sooner than they otherwise would.

The ladder argument is a claim about formation. It says the process of producing that work badly, repeatedly, under correction, is what built the judgement.

Nothing prevents both from holding. Which gives you a specific and unattractive combination:

People produce better work sooner, and understand it less.

Competence arriving ahead of comprehension. The output looks like progress. The formation underneath it may not be happening at all.

I want to be careful here, because this is exactly the shape of argument that flatters experienced people and irritates everyone else. “The young don’t learn properly any more” is not an insight; it’s a genre. The version worth taking seriously has to be more specific than that.

What actually breaks is the proxy

Here is the specific bit.

Organisations have never assessed judgement directly. They assessed output, and used it as a proxy. If the drafts got better, the person was developing. If the drafts stayed weak, they weren’t. Crude, but it correlated well enough to run careers on for decades.

The proxy worked because producing good output required the judgement. You couldn’t fake the draft without having done the thinking.

That’s the link AI severs.

Now output quality and judgement formation can move independently. A first draft can be credible because the workflow is good, the context is rich and the model is capable — none of which tells you whether the person briefing it could have recognised a weak argument, or spotted that the brief itself was wrong.

So the honest problem isn’t that juniors can’t produce. They demonstrably can, sooner than before.

It’s that nobody can tell any more, from the work, whether they’re learning.

That’s a measurement failure before it’s a training failure, and the measurement failure will be discovered second — probably at the point somebody senior leaves and the person expected to replace them turns out to have four years of good output and eighteen months of judgement.

MIT has now described the same break in education

On 25 August, an MIT committee reported that current AI systems can “produce credible solutions and provide reasonable responses to almost any written assignment” in its undergraduate curriculum. The recommended response was not better AI detection. It was to reconsider how learning is assessed, including oral exams, portfolios and in-person conversations tied to work completed elsewhere. MIT’s president described it as a watershed for higher education.

The headline version is that AI can now do the assignments. The more important finding is that the assignment has stopped functioning as reliable evidence of learning. A credible solution may show that a student understands the subject. It may show that they know how to direct a capable system. From the finished work alone, the institution can no longer tell which.

That does not prove students are learning less. The report is an institutional diagnosis, not a longitudinal measure of what students understand. But it is external evidence that the proxy failure above is not peculiar to editorial work or employment. The same separation between output and formation is already visible before people enter the workforce.

MIT’s proposed alternatives all have something in common: they make reasoning more inspectable. That is close to the workplace response I reach for later in the training gap: ask people to show the decisions, checks and changes behind the deliverable, rather than assuming a strong deliverable proves the judgement exists.

The market evidence has moved closer to the mechanism

The project evidence earlier in this piece was internal — one project, one observation, weighted more heavily than its sample size deserved because it was the only thing I had. MIT’s report, above, was the first external evidence to arrive, and it supported the proxy argument rather than the disagreement.

Then came labour-market evidence showing that entry-level roles were becoming scarcer. UK graduate vacancies fell sharply, and SignalFire reported a large decline in new role starts for people with under a year of post-graduate experience across functions including sales, marketing, engineering, recruiting, operations, design, finance and legal.

That mattered, but it still left the causal question open. Falling graduate hiring is consistent with AI substitution, but it is also consistent with a post-2021 hiring correction, ordinary cost-cutting or weaker demand. I wrote at the time that I wanted role-starts and task-composition data before treating the AI explanation as more than plausible.

That evidence has now moved closer to the mechanism.

A new working paper from Bharat Chandar and Bouke Klein Teeselink, How Does AI Change Labor Demand? Evidence from 41 Countries, analyses 1.25 billion job postings and 154 million employment records. The researchers infer firm AI adoption from job advertisements involving generative AI, match foreign affiliates of adopting companies with comparable non-adopting affiliates, and use adoption among nearby multinational parents as an instrument to help separate adoption from selection.

By March 2026, the junior share of employment at treated affiliates was 1.9 percentage points lower than at controls. Senior employment was 6.7% higher. Junior employment was estimated to be 2.5% lower, although that estimate was not statistically significant. Total employment was 3.3% higher, a marginally significant result.

That combination matters more to this argument than a headline about jobs being destroyed.

The firm can grow while becoming harder to enter.

The paper’s computer and mathematical occupations result makes the point even more clearly. These occupations increased their share of employment by 0.8 percentage points in AI-adopting firms, while their junior share fell by 3.3 points. In other words, demand for the occupation can rise at the same time as demand shifts toward people who already have more experience.

That is very close to the ladder problem I have been trying to describe. AI does not need to eliminate an occupation for the apprenticeship route into it to weaken. It can expand the valuable work while reducing demand for the people who used to learn by doing its junior version.

The pattern also travels beyond one country. The researchers could produce country-specific estimates for 31 countries; the junior share fell in 23 of them, with statistically significant declines in seven, including the United States, Brazil and Saudi Arabia.

There are important limits. This is a working paper, not a final verdict. The junior-employment decline itself is not statistically significant, the identification strategy depends on assumptions the authors test but cannot prove completely, and firm-level employment gains do not automatically translate into aggregate employment gains across the economy.

But it changes the balance of evidence again. Earlier, I had evidence that fewer junior roles existed. Now there is evidence, within AI-adopting firms, of a seniority shift that can coexist with modest overall growth.

That is the uncomfortable version of the argument: AI can increase demand for the work while reducing demand for the people who used to learn by doing its simpler form.

What I don’t know

Quite a lot, and it matters here more than usual.

I don’t know whether judgement now forms differently rather than not at all. Reviewing AI output is a real skill, closer to editing than writing, and it’s plausible that it teaches faster than producing from scratch ever did — you see more arguments, more often, and have to decide about each one. If that’s true, the apprenticeship hasn’t disappeared, it’s changed shape and nobody has named it. I’d want that to be true. Wanting it is a reason to distrust my own reading of it.

I don’t know whether anyone is deliberately preserving production work as training, or what it costs them.

And I don’t know how much of my ladder argument is observation and how much is the ordinary human tendency to believe that the way you learned something was the way it had to be learned. That tendency is very strong and almost never examined by the person holding it.

The single observation I do have — that the workflow helped the least experienced person most — is the one piece of evidence pointing away from my own position, which is why I’ve given it more space here than its sample size deserves.

The part that isn’t about content

I should be honest about my own position, because it is comfortable and that is relevant to how much weight this argument deserves.

I am not especially worried about my own job, or about the direction the work I do is heading in. I think I have read that direction right — or right enough for now. Editorial value looks to me like it is moving upstream, into evidence, judgement and narrative, and that is where I have spent the last few years going. Fifteen years of accumulated judgement, a system built to hold it, and a technology that has been straightforwardly good for me.

I hold that loosely. The longer I look at any of this, the less certain I get about the parts that seemed obvious a month ago.

So this is not a piece written from anxiety about my own position. It is the opposite, and that is precisely why the ladder question bothers me rather than reassures me.

My son is starting secondary school. By the time he is choosing what to train for, the entry-level rung of whatever he picks may have been absorbed — and not only in editorial work. Junior legal review. First-year audit. Routine diagnostic coding. Draft analysis. Basic drafting of almost any kind.

The pattern is not specific to my industry. It is specific to work where competence was built by doing the simple version of the complex thing, many times, under correction.

I have no policy view about this, and I am suspicious of anyone who already has one. But I notice that almost every confident account of AI and work is written by somebody who already holds the experience the technology is now amplifying. Including this one.

The question worth asking instead

Not: can early-career people still produce good work? They can, faster than before.

But: if production was doing the teaching, and production has been absorbed, who is designing the replacement?

When I first wrote this piece, my answer was: nobody, as far as I could tell. That is no longer quite true. EY is now an early counterexample: its new Career Residency deliberately combines client work, coaching, professional judgement, critical thinking and AI fluency as the traditional apprenticeship changes. I’ve followed that evidence in The work AI removes was also how people learned.

That does not solve the problem. It makes it more useful. If organisations are starting to design replacements deliberately, we can stop arguing only about whether the ladder is disappearing and start examining which new ladders actually work.

If you manage a team: how are the people who joined recently learning the judgement part, and how would you know if they weren’t? If you joined recently: where are you actually getting it? I’d rather have an answer to either of those than another argument about whether AI can write.

And if you work somewhere that isn’t content or marketing at all — law, finance, engineering, medicine, any trade with a first year that used to teach you something — I would like to know whether this reads as familiar or as somebody else’s problem.