Junior software developer

Systems software developers rank 7th by employment among the highest-exposure quintile in the paper's appendix, and software development is one of two occupations the authors single out for a case study. US private-sector payroll records from ADP covering millions of workers, November 2022 to June 2026. The 11% fall is for 22-25-year-olds in the two most AI-exposed quintiles; the same age group in the three least-exposed quintiles grew about 10%. Experienced workers show no comparable gap, and the authors state they find no evidence of widespread, economy-wide displacement. The findings are descriptive, not causal, and are measured at occupation level, not task level.

Verifiable change in hiring, headcount, hours or job scope. Highest weight — but causal attribution still has to be argued, not assumed.

A census under a statutory duty: 20 fixed questions put to every reporting agency, each answered Y or N with the product named and the licences given as a band. Generating code using AI is Y at 31 of 45, and the products named are ones that did not exist three years ago. A Y is the employer reporting that it has the tool, not a measurement of how much code the tool writes or how much of it is kept, and a licence band is not a count of people writing code with it. The question does not distinguish an engineer of ten years from one of ten months; this record attaches here because writing routine code — CRUD endpoints, forms, standard integrations, tests for known patterns — is what this task is, not because the inventory says anything about junior developers.

An employer has put it into production. Can move the baseline — weighted by scale and how similar the setting is.

An engineering standard binding the Home Office's software teams. It says teams should use AI where it helps, that AI-assisted outputs must be reviewed and approved by a human before reaching production and must be traceable, and that teams will retain full accountability for all AI-assisted code and outputs because AI tools cannot replace human judgement, understanding, ownership or responsibility for decisions, designs or changes made to systems. The firm 'must' applies to human review; the sentence on designs is a statement of accountability. It binds one government department's engineers.

Failure, rollback, regulation or cost is suppressing adoption. Can lower an assessment or widen its uncertainty.

A university study of language models drafting requirements specifications on one project at an IT consulting company, assessed by the company's expert analyst. A subset of user stories stayed unaddressed across all models because the knowledge analysts hold can be tacit and not written in the elicitation documents; the analyst estimated drafting time savings of 10% to 15%, and the authors conclude the models can help early documentation but cannot replace human intervention. It is one project with one expert, and the role studied is requirements analyst rather than developer.

Small-scale trial in a real setting. Tells us the deployment conditions are being tested, not that they hold — so one pilot is never enough on its own; two independent ones are.

One large technology company, company-reported, with no definition of how 'generated' is measured (autocomplete versus whole functions). Says nothing about hiring; the review step stayed with engineers.

An employer has put it into production. Can move the baseline — weighted by scale and how similar the setting is.

Google engineers describing their own internal tooling. Google randomly assigned 50% of all its developers to receive ML-suggested fixes for broken builds in the IDE and compared them with the rest for 11 weeks; afterwards about 34% of users experiencing a build breakage in a month applied such fixes, and ML-powered build repair was enabled for all Google developers. A developer previews each fix and chooses to accept or reject it. These are build and compile errors — the kind the post calls well understood — not runtime faults whose cause is far from the symptom; Google sells developer AI tools.

An employer has put it into production. Can move the baseline — weighted by scale and how similar the setting is.

This occupation has the sharpest split between entry-level and senior of anything in the first batch. The tasks that made junior roles economically worth creating are the ones being automated, while the tasks that define senior work are not. The risk is not that programming stops being a career — it is that the ladder's bottom rung is being removed while the top is intact. Concretely: you need to demonstrate judgement, not throughput, and you need to do it before anyone would normally expect it of you.

Your position is comparatively strong, but it depends on a supply of people who learned judgement the slow way. If the entry rung stays broken, the constraint arriving in a few years is not automation — it is that nobody built the experience to review what the tools produce. Worth thinking about as a hiring and mentoring question now rather than later.

Stay and strengthen

Build judgement faster than the ladder expects

The scarce thing is someone who can tell whether generated code is right. You can start demonstrating that on day one instead of waiting two years.

Reshape the role

Specialise where correctness is expensive

Payments, security, data integrity, infrastructure — domains where a confident wrong answer costs real money keep humans firmly in the loop.

Adjacent move

Roles that sit between engineering and the problem

Solutions engineering, developer experience, technical product — jobs where the value is understanding both the system and what someone needs from it.

Cross-industry

Apply engineering thinking in a non-software domain

Many sectors have almost no one who can both understand their domain and build. That combination is scarce and does not compete with generation.

Yes, but for a different reason than five years ago. Learning to code used to be about being able to produce code. Now the value is in being able to judge it — and you cannot judge code you could not have written. The training still works; what has changed is that the entry-level job market no longer reliably pays you while you acquire the judgement. Plan for that gap explicitly rather than assuming the first job will teach you.

The economic reason those roles existed — cheap throughput on well-specified work — is weakening, and that is a real structural change rather than a cycle. But teams that stop hiring juniors entirely create a problem for themselves: in a few years they have nobody who learned to judge output the hard way. Expect the role to be redefined around review and judgement rather than production, and fewer of them per team than before.

The signal for entry-level developers is not model capability, it is your employer's graduate intake. Look at how many juniors your company hired this year versus two years ago. Several organisations that cut junior intake have publicly said they removed the path that produces seniors — that admission is the event worth tracking.

Writing routine code — the task marked automating, and the one that used to fill a junior's first two years. System design and requirements are not, but they are where you arrive after those two years, and nothing has replaced the road.

Yes, and for a reason that has changed: you now learn to code in order to review code, not to produce it. Reviewing fluent-but-wrong output well requires having written the same thing yourself, which means the learning still has to happen the slow way even though the market for the output has shrunk.