In ADP payroll records, employment of US software developers aged 22 to 25 is down almost a fifth since November 2022, the month ChatGPT launched and close to the young group's peak: its index stands at 81.4. Ages 26 to 30 are down 4%. Every group over 30 grew; ages 41 to 49 are up a fifth, at 120.2. Public survey estimates are volatile; this analysis does not establish whether the age pattern is statistically distinguishable from sampling variation.
Three releases in the past month bring the picture up to date: Stanford's payroll index on 16 October, with data to September; Indeed's job ads, to 30 October; and the New York Fed's figures for recent graduates on 2 November.
What these numbers are
The chart counts people in the software developer occupation at a five-year balanced sample of firms that use ADP, a large US payroll company: the same firms throughout, not national employment. Erik Brynjolfsson, Bharat Chandar and Ruyu Chen at Stanford built it for their paper "Canaries in the Coal Mine?" and update the index monthly on a public dashboard. Each release can revise past months.
It is an indicator of young-worker employment, not a direct measure of hiring or seniority. Stanford's base month, November 2022, sits next to the young group's peak. Measured from October 2021 instead, the first month in the latest release, 22–25 is down 9.7% and 41–49 up 30.7%. Ages 22–25 are the only group below their starting level.
Ads fell first
Indeed's index counts job ads, not hires. Ads for software developers peaked in February 2022 at 2.3 times their pre-pandemic level, then fell by almost three-quarters to a low in May 2025. On 30 October 2026 they stood at 81.6, 18% below February 2020, while all postings stood at 103.8.
About 60% of the net decline from the sampled peak to October 2026 had occurred by 26 November 2022, before ChatGPT launched. In this post-peak slump, postings first fell below their pre-pandemic baseline in April 2023. Timing alone does not identify AI's contribution.
In all six countries for which Indeed publishes software data, developer ads fell by between 63% and 71% from their peak, more than ads overall. The US drop, 65%, is close to Britain's and Canada's and smaller than Germany's.
US software postings are up a third from their May 2025 low. The youngest developers have not shared in it: their payroll index is lower than in May 2025. Indeed's economists find 71% of the increase in software postings between May 2025 and May 2026 was for senior roles, and in early 2026 just 4.5% of software postings were entry-level.
What the household survey shows
The household survey provides a small-sample cross-check. Its share of software developers aged 22 to 25 was 7.1% in the year to September 2026, the lowest in six years, but it was 9.1% the year before. Estimated employment of developers aged 22–25 was about 174,000 per month, above 159,000 in 2021 but below 201,000 in 2025. These small-sample estimates are noisy. The 2026 window contains 11 observed months; October 2025 was not collected.
The Stanford authors note their pattern is "more pronounced in the ADP analysis sample than in national survey benchmarks". The strongest age evidence comes from one payroll company's clients.
Recent computer science graduates had an unemployment rate of 7.0% in 2024, against 4.2% for all recent graduates (New York Fed; graduates aged 22 to 27 not enrolled in school). In the New York Fed's 2 November update, the rate for all recent graduates was 5.2% in September 2026, against 4.0% for workers aged 16–65 not enrolled in school (seasonally adjusted three-month averages).
What might explain it
The data shows when and to whom, not why.
- AI coding tools. The Stanford authors find young workers in AI-exposed jobs well behind their peers, but call their results descriptive, "rather than causal estimates"; with their strictest controls, the decline in AI-exposed jobs becomes significant only in 2024. Hosseini and Lichtinger find junior employment falls at firms that adopt generative AI. In October 2025 the Yale Budget Lab found no "discernible disruption" in the wider labour market.
- Over-hiring and interest rates. Iscenko and Curto Millet of the Economic Innovation Group blame "the sharpest monetary policy tightening cycle in four decades". Software postings peaked on 26 February 2022, less than three weeks before the Fed's first rise. Morgan Frank and colleagues find AI-exposed job markets were deteriorating from early 2022, before ChatGPT. The Stanford authors reply that AI-exposed jobs are, if anything, less sensitive to interest rates.
- Section 174. From 2022 US firms had to spread domestic software development costs over five years instead of deducting them at once, and foreign research costs over 15 years (Congressional Research Service), until a 2025 law restored immediate deduction for work done in the US (IRS). Evidence that it cut hiring is anecdotal. This descriptive comparison cannot isolate a US tax-policy effect.
- Offshoring. We found no systematic study either way.
One company has published its own numbers. On 16 October, the AI company Anthropic reported a similar change in hiring composition: junior roles fell from 27% to 9% of its technical hires since early 2025. This does not establish an industry-wide causal effect. Researchers quoted by the Guardian called it an unaudited account from a company that could list its shares as soon as November.
How this was made
Four sources, snapshotted on 7 November 2026: Indeed Hiring Lab's job postings data (CC BY 4.0), Stanford's index, Census survey microdata for every available month from January 2020 to September 2026, excluding October 2025, (aggregated by fetch-cps.mjs; October's file is not yet out) and the New York Fed's graduate data. The extracts are here: Indeed, Census survey, New York Fed by major and over time.
Stanford's index has no stated licence, so it is charted and cited here. The monthly series is not redistributed; selected summary statistics are included. fetch-canaries.mjs downloads it from Stanford; the manifest records the hash of the 16 October 2026 release used here, and later releases revise past months.
calculate.mjs computes the snapshot-derived statistics; junior-dev-hiring-make-charts.mjs builds the charts. Licences and choices: PROVENANCE.md and DATA.md. From the download folder, run:
npm ci
node junior-dev-hiring-src/fetch-canaries.mjs
node junior-dev-hiring-src/calculate.mjs
node junior-dev-hiring-make-charts.mjs
Claude, an Anthropic model, assisted with drafting; sources and calculations were independently reviewed by Claude and OpenAI models.