Research update · 11 October 2026
Copilot studies split on hiring. They count different things.
A 2026 NBER conference abstract estimates about 35% fewer technical-worker hires within a year among large firms induced to adopt GitHub Copilot by a contract renewal. A separate April journal article links Copilot adoption to a 3–5% higher monthly chance of hiring a software engineer. Neither number describes every employer.
What the 35% estimate counts
Dillon and Raymond describe 5,425 large multinational firms employing 45 million people worldwide. Their conference abstract says pre-existing enterprise software contract-renewal cycles shifted Copilot adoption timing. Among firms induced to adopt by that timing, technical-worker hiring fell about 35% within one year. The authors report no matching increase in separations; code-intensive occupations saw the largest change, and hiring in the most code-intensive role shifted toward senior workers.
This is an author-reported estimate for a selected adoption margin, not a claim that 35% of programmers lost jobs. The full paper and slides were not accessible in this run, so precision, the comparison group and robustness checks remain unverified.
What the positive estimate counts
Baird and colleagues use LinkedIn and GitHub records and report a 3–5% higher monthly probability of hiring a software engineer at firms adopting Copilot, mainly through entry-level hiring. The paper calls this an association. A higher chance that a firm hires at least one person in a month can coexist with fewer total hires at a different set of large firms. It is not a 3–5% increase in headcount or hiring volume.
The journal article's abstract was accessible, not its full text. The underlying data are proprietary. Three authors work at LinkedIn and one at GitHub; those links do not invalidate the result but make independent replication important.
What changed in the answer
We now have a sharper adoption-linked hiring warning for technical work and an opposite-direction association for software-engineer hiring incidence. They are not a clean contradiction: firm samples, timing, comparisons and outcomes differ. Earlier junior software-ad data measured vacancies, not hires. The three measures must stay separate.
A U.S. worker survey also reports that AI use is widespread but uneven across tasks. Use of a coding assistant is not itself a layoff. The next useful test would put the same firms, dates and adoption definition against both the number of hires and the chance of any hire, then check separations, junior share and pay.