Chapter 20 — Growing Juniors When the Bottom Rungs Are Automated

Nobody designed junior work as a curriculum. It was a curriculum by accident, and the accident has stopped.

20.1 The seventh good

Chapter 5 argued that six goods were purchased by the expense of writing code — scrutiny, intent, pacing, a trust boundary, structural homeostasis, and the premise that production was the expensive step — and that all six failed together when the price collapsed.

It missed one, and the omission matters more than the others because its failure is the slowest to appear.

Consider what junior work actually consisted of: fix a small bug, add a field, write a missing test, update a dependency, follow an existing pattern in a new place. And ask why those tasks went to juniors.

Not because they were pedagogically designed. They went to juniors because a senior's time was expensive and the task was simple. Pure labour arbitrage. The allocation was economic, and the curriculum was accidental.

But look at what the accident taught.

Now remove the arbitrage. A small bug costs nothing to fix, so there is no economic reason to route it to a human at all — let alone to the least experienced one, who will take longest.

Nobody decided to stop training juniors. The mechanism that did it was never named, which is precisely the pattern Chapter 5 identified: goods that arrive free are goods nobody audits, and their withdrawal is invisible until something downstream fails.

Which places this chapter squarely inside the book's own definition. Chapter 5 §5.7 defined ADLC as the discipline of explicitly buying what the cost of writing code used to provide for free. Apprenticeship is the seventh item on that list, and it is the one where the bill arrives last.

20.2 The bifurcation

Two of the best-evidenced findings in this book point in opposite directions on experience, and putting them together produces the chapter's central problem.

Juniors gain more. Cui, Demirer, Jaffe, Musolff, Peng, and Salz pooled three randomised controlled trials across 4,867 developers at Microsoft, Accenture, and a Fortune 100 company, published in Management Science. Overall: +26.08% completed tasks (SE 10.3%). And specifically: less experienced developers had higher adoption rates and greater productivity gains [A].

Seniors gain less, or lose. METR's randomised trial of 16 experienced developers on 246 real tasks in repositories they had worked in for an average of five years found AI-allowed tasks took 19% longer [A].

And juniors are the more exposed to the cognitive cost. Lee et al. found that self-confidence predicts more critical thinking while confidence in the tool predicts less [A]. A junior has, by definition, less warranted self-confidence and more reason to defer. The MIT study's mechanism — shallow encoding, poor recall, work not internalised — bites hardest on someone who has not already consolidated the competence [B]. And Burelli et al.'s reduced-engagement pattern did not differ between undergraduate and graduate students, offering no evidence that the effect attenuates with the (limited) experience range they sampled [B].

This is a trap with three properties, and it is the specific combination that makes it serious:

It is individually rational. A junior who delegates heavily produces more, faster, and looks more capable than one who does not.

It is maximally rewarded. Under any throughput measure — and Chapter 12's seventh prohibition exists because organisations reach for throughput measures — the heavily-delegating junior is the better performer.

And it is invisible to the person in it. Chapter 18 established that under AI assistance the internal signals which track performance decouple from it. Chapter 19 established that the faculty being degraded is the one that would assess the degradation. A junior in this trap has no reliable internal instrument for noticing, and the external instruments that would tell them do not exist in most organisations.

20.3 The employment signal

The labour-market evidence is stronger than the version circulating in secondary coverage, and more carefully hedged by its own authors.

Brynjolfsson, Chandar, and Chen, at the Stanford Digital Economy Lab, analysed high-frequency administrative payroll microdata from the largest payroll software provider in the United States — administrative records rather than a survey [B]. Their paper is titled Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence, and the question mark is the authors'.

• Early-career workers aged 22–25 in the most AI-exposed occupations show a 13% relative decline in employment, after controlling for firm-level shocks.
• A later update reports the gap widened: employment of 22–25-year-olds in AI-exposed occupations stands 19% below where it would be had it kept pace with less-exposed peers.
• Experienced workers show no comparable gap.
• Software engineering is among the most exposed occupations named.

And the finding that keeps this honest, which most coverage omits: the researchers do not observe widespread, economy-wide displacement. Employment in less-exposed fields, and for more experienced workers in the same occupations, has remained stable or continued to grow.

So the correct reading is not "AI is destroying software jobs." It is narrower and more precisely aimed at this chapter's subject:

Within exposed occupations, there is an age gradient. The entry-level rung is where the effect concentrates, and experienced workers in the very same jobs show nothing comparable.

Three cautions I would rather state than have quoted back.

The design controls for firm-level shocks, which is strong — it compares within firms, so macroeconomic conditions and sector-wide layoffs are substantially handled. But "AI-exposed occupation" is a classification, not a treatment, so this remains observational.

The authors' own framing is canaries — an early indicator warranting attention, not a settled causal finding. Their title carries a question mark and this book should carry it too.

And the ledger's earlier version of this claim, drawn from secondary coverage, said "~14% slowdown in hiring." That is a different quantity from a 13% relative employment decline, and the corrected figures are the ones above. A second circulating figure — that older developers' employment grew 6–12% in the same firms — is not verified here and is not printed.

20.4 Why this is not a training problem

The instinctive organisational response is a training programme: teach juniors to use AI well, add a module on prompt engineering, run a workshop.

Chapter 17 §17.5 explains why that fails. The shifts divide into informational ones, which respond to being told, and behavioural ones, which do not. Everything at stake here is on the second list. METR's cohort knew the result and still felt fast. A workshop transmits information to a faculty that is not the one making the error.

There is a deeper reason, and it follows from §20.1. The problem is not that juniors lack knowledge about how to work with agents. It is that the experiences which used to build judgment have been removed from their week. You cannot restore an experience by describing it. The gap is in hours spent being wrong about a real system and finding out, and no curriculum delivers that by exposition.

So the framing that follows from the book's own thesis: this is not a training problem, it is a work-allocation problem. Chapter 5's remedy for every withdrawn good was to buy it deliberately. The purchase here is deliberately allocated work whose purpose is learning rather than delivery — which costs delivery, which is why it will not happen by default.

20.5 The ladder as a curriculum

Chapter 13 built the autonomy ladder as a governance instrument: A0 through A4, with the level set by harnessability × blast radius and never by enthusiasm. Read it a second way and it is a developmental sequence.

This reframing has a practical consequence that runs against current practice. Organisations currently place juniors at A2 immediately — hand them an agent, let them complete scoped tasks, review the diff — because it maximises their output, which Cui et al. confirms is genuinely higher.

Chapter 13 already noted that A0 is the correct level for some work permanently. The addition here: A0 is also the correct level for some people, temporarily, and for reasons that have nothing to do with the blast radius of the change. Assigning a junior to author something by hand is not a statement about the code. It is a statement about what the hour is for.

Which requires saying out loud what the hour is for, because otherwise it reads as distrust. The ladder gives you the vocabulary to make that explicit: this task is A0 for developmental reasons, not risk reasons. Naming it converts an insult into a curriculum.

20.6 Designed practice

If learning is no longer a by-product, it has to be designed. Four practices, each derived from a verified finding rather than from how apprenticeship used to feel.

One: attempt-first is mandatory, not optional. Chapter 19 §19.7 reported the ordering effect — participants who worked unaided first and used the tool afterwards showed higher recall and preserved prefrontal activation, while the reverse ordering produced shallow encoding [B]. For an experienced engineer, §19.7 recommended choosing which competences to protect and paying the minutes on those. For a junior, there is no established competence to triage against, so the rule is unconditional: form your own answer first, even briefly, even badly, and only then delegate. The cost is minutes; the difference is whether the task deposits anything.

Two: the prediction log is a training instrument. Chapter 18 §18.7 proposed recording will this need changes? before opening a diff, as a calibration device. For a junior it is more than that — the hit rate is the competence measure, and it is one of the very few external instruments that reports on a faculty whose internal signal Chapter 18 showed is decoupled. A junior with forty logged predictions has something no amount of self-assessment provides: evidence about their own judgment.

Three: verification practice requires ground truth, which means their misses must be found. You cannot learn verification judgment without discovering where you were wrong. This is where Chapter 16's sampling audit — deep re-review of a random sample of already-approved changes — earns a second purpose. Proposed there for measuring organisational escape rate, it is, pointed at an individual, an apprenticeship instrument: the only systematic way a junior learns what they let through.

The statistical limits of §16.7 apply and are less damaging here. That section noted a 5% sample at a 2% escape rate yields fractional findings and cannot give a team a usable quarterly number. But for pedagogy you do not need an unbiased rate estimate — you need instances to learn from, and any miss found is a lesson regardless of what it implies about the population.

Four: build self-efficacy, not tool-scepticism. Lee et al. identifies self-confidence as the protective variable, and Chapter 18 showed the other lever is already pulled — developers report historically low trust and it does not protect them. So the intervention is anything that produces genuine, demonstrated competence, and the prediction log and audit above are exactly that, because both produce evidence rather than encouragement.

20.7 Apprenticeship competes with the constraint

Now the hard part, which most writing on this subject omits.

The traditional apprenticeship channel was review: a senior reads a junior's work carefully, explains what is wrong, and the explanation is the teaching. It worked because review was already happening and the pedagogy was, again, a by-product.

Chapter 3 established that review capacity is the binding constraint. Chapter 16 reported the enterprise measurement: per-reviewer load roughly doubled while per-capita throughput rose 2.09× [B].

So apprenticeship-by-review now competes directly for the scarcest resource in the organisation, against throughput, at exactly the moment that resource came under maximum pressure.

There is no clever resolution. Apprenticeship review and throughput review are different activities and cannot be done with the same hours. A senior reading a junior's change to teach is reading slowly, asking questions whose answers they already know, and explaining. A senior clearing a queue is doing none of that.

Which means it has to be separately budgeted, and that is Chapter 16's public-good argument arriving one more level down: the cost is borne by a team this quarter and the benefit accrues to the organisation in three years, to a person who may by then work somewhere else. No team will fund it, and none should be asked to.

20.8 What organisations owe the pipeline

The structural claim, stated as plainly as I can.

Demand for senior judgment is rising at exactly the moment the mechanism that produces seniors is being removed. Chapter 8 established that the failure classes neither computational nor inferential sensors reliably catch — misdiagnosis, over-engineering, unrequested features, correctness against unstated intent — are the residue that requires human judgment. Chapter 19 established that specification judgment is the faculty the lifecycle now depends on most. Chapter 15 established that no amount of harness maturity substitutes for either.

Seniors are made, not born, and they were made on the rungs that no longer exist.

Three things follow for anyone running an engineering organisation.

Junior work must now be chosen rather than inherited. The tasks that used to fall to juniors because they were cheap will not fall to them any more. Someone has to decide, deliberately, which work a person does by hand and why — and Chapter 13's ladder is the vocabulary for saying so without it reading as demotion.

The teaching budget must be explicit. Figure 20.4 shows what happens to an implicit one. A named allocation of senior hours to apprenticeship review, protected from throughput pressure, is the minimum, and it will lose every prioritisation argument it is entered into unless it is protected in advance.

And the measurement system must stop penalising it. Chapter 19 §19.8 made this point for individuals maintaining their own competence; it applies with more force to seniors spending hours teaching. Under a throughput mandate, the engineer who teaches is underperforming, measurably, on the metric being watched. That is Chapter 16's Figure 16.1 again, and it is now the third time the same structure has appeared in this book — which is a reasonable sign it is the actual shape of the problem.

One honest counter-argument, and it is the strongest one against this chapter.

Perhaps juniors trained in this environment simply develop different competences — verification-first rather than authorship-first — and those competences are adequate or better for the work that will actually exist. This is possible. It is untested. And the specific reason for concern is narrower than nostalgia: Chapter 19 showed the faculty they would be exercising, verification judgment, is the one whose feedback signal has been decoupled from performance. Learning by doing requires the doing to inform you. If it does not, practice may not converge on competence — and that is an empirical question nobody has answered, about a cohort that is being run through the experiment now.