The cleanest public figure is 54,836: the number of US layoffs in 2025 that employers explicitly attributed to AI [E1]. The tally records corporate attribution, not a counterfactual measure of jobs that would have survived without the technology [E1]. Tech companies reported about 152,922 layoffs in 2025 and 120,846 more through 13 July 2026, leaving the AI-cited slice material but smaller than the broader retrenchment [E2]. The denominator blocks both easy stories, that AI explains every cut or none of them [E1][E2].
CrowdStrike supplied the rare clean disclosure: a May 2025 filing paired an approximately 500-job cut with the statement that AI “flattens our hiring curve” [E3]. Amazon’s chief executive separately told staff that generative AI and agents meant “fewer people doing some of the jobs” and would reduce the company’s total corporate workforce [E4]. Oracle said in its annual report that AI adoption had “resulted” in workforce reductions as headcount fell by about 21,000 over the year [E5]. These cases connect the tool, the workforce effect and the executive rationale in the same record, which makes them stronger evidence than retrospective branding [E3][E4][E5].
Most layoff announcements do not offer that chain [E1]. Executives can deploy useful automation, remove management layers, trim pandemic-era hiring and finance a larger AI infrastructure bill in the same restructuring, then give investors the shortest explanation [E1]. Challenger’s analysts warned that AI’s true effect is hard to size because leaders are rewarded merely for naming it [E1]. AI therefore functions in three roles at once: operating technology, budget justification and market signal [E1][E3][E4].
The coding layer is where the mechanism becomes visible. By May 2026, Claude had authored more than 80% of code merged into Anthropic’s own codebase, up from low single digits a year earlier [E6]. OpenAI said its average engineer generated 99% of output tokens through Codex, while Dario Amodei said coding would disappear before software engineering [E7][E8]. Those figures describe tools absorbing codified, lower-context tasks at the point where companies once trained junior engineers through repetition [E6][E7][E8].
Hiring data shows the first pressure at the bottom of the ladder. New-graduate hiring was about 65% below 2019 levels at the largest technology companies and about 76% lower at early-stage startups [E9]. Workers aged 22–25 in AI-exposed occupations recorded a roughly 16% relative employment decline, a sharper signal than the aggregate layoff totals [E9]. The New York Fed, however, found little distinct AI-driven collapse in overall job postings, so the evidence supports concentrated entry-level damage before an economy-wide jobs shock [E9].
Klarna supplied the clearest warning against treating a demonstration as a settled labor model. In 2024 it said its assistant performed the work of 700 customer-service agents [E10]. During 2025 the company resumed hiring human support, and its chief executive stressed that customers must always be able to reach a person [E10]. The reversal did not erase the automation; it exposed the service, exception-handling and trust work that the headline number had compressed [E10].
The honest reading is untidy because the corporate process is untidy. Automation removes some tasks, executives use AI to sell conventional restructuring, and investors reward a story of labor efficiency while companies keep paying for compute and reorganization [E1][E3][E4]. The aggregate numbers remain too modest to prove a general employment collapse, yet the coding data and junior-hiring decline show a real apprenticeship squeeze inside software work [E2][E6][E7][E9]. The machine can perform the junior task today; the company still has to explain who becomes senior tomorrow [E6][E7][E9].