Two things happened to the junior developer market this year, and only one of them gets reported.
The first is a number. Junior software developer employment has fallen for 33 consecutive months. It turns up in every piece about the end of the entry-level job, usually in the first paragraph, usually with a link to Stanford attached.
The second is that in February, IBM said it would triple its entry-level hiring in the US this year. Its HR chief added a line that should have travelled further than it did: "And yes, it's for all these jobs that we're being told AI can do."
I went looking for the 33 months. It is not where everyone says it is.
The streak the paper never counted
The trail runs to Stanford's Digital Economy Lab and its Canaries in the Coal Mine? paper, updated in August 2026 with ADP payroll data through June. It is good work: administrative records covering millions of workers, published on a rolling basis, and unusually honest about its own limits.
Search the paper for the streak and it is not there. No "33 consecutive months", no "October 2023", no year-over-year streak of any kind. The figure gets assembled downstream from the lab's public dashboard, then labelled as something the paper never claimed.
Two things get lost in the assembly.
The label says software developers. The measurement is workers aged 22 to 25 in AI-exposed occupations, which is a far wider group: customer service representatives, administrative roles, a long tail of codified knowledge work.
The authors go out of their way to say so. Drop computer occupations from the sample and the estimate is "essentially unchanged". Drop technology firms and it is only slightly smaller. Their conclusion, in their words: "our findings are not specific to technology roles."
The second loss is statistical. If you want a monthly series for young software developers specifically, the public data will not carry it. The paper's own appendix reports that the Current Population Survey contains "between 23 and 48 young software developers aged between 22 and 25 per month", with "common fluctuations of 20% or greater in estimated employment month-to-month". A 33-month streak counted on a sample of thirty people is measuring the sample.
None of which makes the decline fake. Employment of 22 to 25 year olds in the two most AI-exposed quintiles fell about 11% between November 2022 and June 2026, while the same age group in the three least-exposed quintiles grew about 10%. Against a kept-pace baseline the gap is 19%, widened from 15% at the previous year's data vintage. That is a real divergence and it is still opening.
The paper's headline finding is the one nobody puts in a headline: "no evidence of widespread, economy-wide job displacement from AI."
The two facts underneath it
The paper numbers its findings, and the fourth should change the conversation. The divergence "operates primarily through reduced hiring rather than increased separations."
The junior was never hired in the first place. Nobody had to be replaced.
That is a purchase order that stopped being signed, quarter after quarter, by people who each had a defensible reason for their own signature. I made that case from the demand side when the resilience headlines came out, and the payroll data now puts a mechanism under it: an industry-wide hiring pause wearing the costume of technological inevitability.
The fifth is the other half. The declines concentrate in occupations where AI is used to automate the work. Where it is used to complement the worker, employment is flat or rising. The classification comes from the Anthropic Economic Index, mapped onto occupations.
So the finding is narrower and more useful than the headline it produced. Wherever the job had been defined as the part a model can do alone, the job stopped being offered.
Which raises the obvious question. What happens when somebody redefines the job?
The company that decided the other way
IBM did, in February. Tripling US entry-level hiring in 2026, across departments, in the same year its peers were explaining thinner graduate intakes with the word "AI".
Nickle LaMoreaux, IBM's chief human resources officer, described the redesign plainly. She changed the descriptions for entry-level jobs so they were "less focused on areas AI can actually automate, like coding, and more focused on people-forward areas like engaging with customers".
Her reasoning is the seed-corn argument arriving on a balance sheet. Cutting early-career hiring saves money now and produces a shortage of mid-level managers later, at which point you buy them from competitors, slower and at a premium.
IBM left something out of the announcement. The company declined to give a baseline, so "triple" is a ratio and nothing more. Three times a gutted 2025 intake can still be smaller than 2019.
IBM is also running rolling workforce reductions through 2026, at levels its own January earnings call described as fairly consistent with the prior year. This is a company changing the shape of its workforce, buying at the bottom while cutting the middle, and the entry-level expansion is one line in that trade.
The reasoning still matters. IBM is the only large employer publicly betting against a consensus its peers describe as physics.
Where the bet is right
Put IBM's redesign next to Stanford's fifth fact and they line up almost exactly. The declines land where AI automates. IBM moved the junior role off the automated tasks and onto complementary ones. That is the correct direction, arrived at independently, and it is more than anyone else has done with the same data.
Where it might not pay
The paper has one more result, buried in a regression table, and it is what I would put in front of anyone about to copy IBM.
Automation and complementarity were entered jointly against employment change, by age group. For 22 to 25 year olds, only the automation coefficient is statistically significant, and it is negative. The complementarity coefficient is positive and significant for workers aged 41 to 49, and for those over 50.
Complementarity has been paying off for people who already had something to complement.
That is the thing IBM's bet has to overcome. A complementary role assumes judgement already in the building. The junior is the person who does not have it yet, which is the entire reason the role got cut.
The complementary work that actually expanded is also the work a junior can least do. The review queue is where the senior time went: telemetry across four thousand teams puts time to first review up 156.6%, with code arriving that is "often not review-ready", so the reviewer finishes it. Reviewing an agent's output well means knowing what the code should have done, which is the last thing you learn rather than the first.
Then read LaMoreaux's reasoning once more. She said the shortage would be in mid-level managers.
She is probably right, and that is a different shortage from the one the seed corn describes. Customer-facing work with AI oversight produces people who can run a team and handle a client in five years. It does not produce the person who can look at nine hundred generated lines and say which forty are wrong.
Both gaps are real. One of them is being funded.
If you copy the bet, copy more than the headcount
Hiring the junior is the cheap half. The expensive half is what you point them at on Monday.
- Put them in the review queue as a second reader, never the approver. Reading a senior's review of an agent's pull request alongside the diff is the fastest compression of judgement available right now, and it costs the senior almost nothing.
- Give them the small dangerous changes, with a name attached. A migration on a staging table, a permission boundary, a retry that could double-charge someone. Consequence is the teacher, and supervision is what keeps the tuition affordable.
- Count reading, not tickets. A junior whose week is measured in merged pull requests will let the agent write them. One whose week includes explaining someone else's code back to them is turning into the reviewer you are short of.
That is the cheapest senior engineer you will ever buy, on a five-year delivery schedule, which is roughly the lead time on the shortage everyone is currently pricing into their 2031 recruiting budget. I wrote the half of this bargain that a junior controls in what a junior should be learning instead. This is the half that belongs to whoever signs the offer.
The collapse was never a verdict on what a junior is worth. It was a decision, taken separately by thousands of companies, that this year's roadmap mattered more than 2031's engineers. IBM has signed the order everyone else stopped signing. Whether that grows seniors or merely staff depends entirely on what it asks them to do on Monday morning.