What This Is
Stanford HAI, together with three scholars (Brynjolfsson, Chandar, and Chen), released a labor market study. The data comes from anonymized payroll records at US payroll giant ADP, covering 25 million employees, tracking employment changes along two dimensions: age and role.
The core number is 19%—young people aged 22 to 25 working in software development, customer service, accounting/auditing, administrative assistance, and other high-AI-exposure roles (work whose tasks can be substantially taken over by AI tools) show employment levels 19 percentage points lower than peers in low-exposure fields. Last year the gap was 13%; it is widening. By contrast, 35-to-49-year-old senior employees in the same fields see employment actually grow by roughly 9%.
We see the study attributing the difference to "knowledge type": what AI replaces is the bookish, standardized, template-able "codified knowledge"—precisely what newcomers mainly possess at career entry. The "tacit knowledge" that senior employees have accumulated (interpersonal judgment, complex situational handling, cross-functional coordination) AI still cannot learn. What young workers are losing are the entry-level jobs that should have served as their stepping stones.
Industry View
Supporters argue the study made serious causal exclusions—stripping out the tech hiring slowdown, the spread of remote work, and post-pandemic declines in education quality—the 19% gap remains robust and can be treated as strong evidence of "AI structurally displacing entry-level roles."
Criticism exists too. First, the study itself concedes it is "not entirely AI-driven"; macro cycles and industry restructuring may still contribute. Second, the data comes from a US payroll system, so its explanatory power for the Chinese market is limited. Third, although the 19% relative gap is significant, the magnitude of the absolute employment decline still needs watching—we cannot directly extrapolate to "young people cannot find work."
Even more worth our vigilance, in our view, are low-education workers. The study notes that in low-education jobs, the employment gap across different AI-exposure levels can persist all the way to age 40. For those without an educational shield, experience cannot buy protection, and AI's replacement speed may outrun their skill-upgrade speed. This is a more severe, yet less discussed, problem than "newcomers can't find jobs."
Impact on Regular People
For enterprise IT and hiring: the entry-level "learn by doing" mechanism is failing. Companies need to redesign cultivation paths—shifting from "hire them in and teach them slowly" toward denser upfront capability screening, or using project-based internships to replace traditional campus-recruiting filtering logic.
For individual careers: we see the competitive dimensions for the 22-to-25 cohort shifting. "Can write code" and "can recite standard processes" are no longer moats. The abilities to handle ambiguous situations, communicate across departments, and make complex interpersonal judgments are scarcer, and harder for AI to take over.
For consumer markets: short-term impact is limited, but over the medium-to-long term, the actual purchasing power of entry-level white-collar workers may decline. Vocational retraining, AI tool application courses, and cross-domain skill migration services may become new consumer tracks.