The automation wave is no longer coming for the factory floor. It is coming for the office, the campus, and the innovation hub, and the first casualties are the youngest workers in the most AI-exposed occupations. That is the core finding of two major studies covered this month by CNBC’s Deena Zaidi, and it upends the comfortable assumption that AI would augment knowledge workers rather than displace them.
The numbers are stark. The American AI Jobs Risk Index from the Digital Planet initiative at Tufts University ranks 784 U.S. occupations across 20 industry sectors and finds writers and authors at 57% vulnerability, computer programmers at 55%, and web and digital interface designers at 55%. The largest total income loss falls on software developers, management analysts, market research analysts, and marketing specialists, a function of both their high salaries and their sheer numbers. These are not routine clerical roles. These are the jobs that a decade of tech-industry rhetoric promised would be protected by the very skills AI now reproduces.
The paradox at the heart of AI exposure
The Tufts study’s most counterintuitive finding, articulated by professor Bhaskar Chakravorti, dean of global business at Tufts’ Fletcher School, is that the more AI helps you do your job, the more expendable you become. “The parts of the country or the jobs that are most helped by the technology are also the ones that are most hurt by it,” Chakravorti told CNBC. “If you’re in high tech, you are also doing exactly the kind of work that AI is getting better and better at doing.”
This is a direct inversion of the last two decades of tech labor economics. When software ate the world, the people who built the software were the winners. The current wave targets the builders themselves. Coding, summarizing, researching, analyzing, and generating first drafts are precisely the core tasks that large language models have commoditized, and the Tufts study argues that risk does not diminish with the value of the work. A cardiologist’s judgment remains hard to replicate. A junior developer’s first-draft code is exactly what a model produces best.
The young pay first
The Stanford Digital Economy Lab’s research, led by director Erik Brynjolfsson and detailed in the report “Canaries in the Coal Mine,” adds a generational dimension that should worry anyone building AI tools for the enterprise. Using ADP payroll data on millions of workers, the study found that employment for early-career workers aged 22 to 25 in the most AI-exposed occupations had fallen 16% relative to their peers. Older workers in the same occupations are, in Brynjolfsson’s words, “largely holding steady.”
The mechanism is elegant and brutal. “AI is a substitute for book knowledge, which a new grad brings,” Brynjolfsson said. “It’s a complement to tacit knowledge, what experience builds.” A fresh graduate’s value proposition was always the ability to apply recently learned formal knowledge at scale. That is now the cheapest thing a model can do. The experienced worker’s judgment, built over years of context and failure, remains something models cannot yet reproduce.
The counterexample proves the point. In a study of customer service agents published in The Quarterly Journal of Economics, Brynjolfsson and colleagues found that the least experienced workers gained the most from AI assistance, improving productivity by 34% compared to a wider average of 14%. In assistive contexts, where AI augments rather than substitutes, entry-level employment has held up and in some cases grown. The difference is not the technology. It is the design of the workflow around it.
A different kind of wave
Brynjolfsson draws the historical contrast sharply. “Steam engines hit muscle work and earlier software hit routine clerical work,” he told CNBC. “But generative AI helps with many cognitive tasks — writing, coding, analysis — the bread and butter of well-paid knowledge work. That’s new.”
The speed is also new. Industrialization took decades to reshape labor markets, and for a generation ordinary workers’ wages moved slowly while output soared. Brynjolfsson warns that this wave is moving much faster. The labor market effects measured now are, in his words, “the leading edge, not the full wave,” and most workers still barely use these tools. The displacement we are seeing is happening before widespread adoption, which should give every AI builder pause about what full deployment looks like.
Neither study predicts mass overnight unemployment. Brynjolfsson is emphatic that “no job is a single task” and that even the most exposed occupations contain plenty of work AI cannot do. The Tufts study similarly shows physicians, including cardiologists and psychiatrists, appearing less exposed despite high salaries, with Chakravorti describing a degree of augmentation that frees healthcare professionals to serve more patients in the same time. The disruption is real, but it is a reshaping of tasks, not the elimination of job categories.
What this means for AI builders
The policy implications are significant, but the more urgent read is for the industry building these tools. The 16% decline in young worker employment is not an externality. It is a product feature. Every enterprise AI tool that automates first-draft writing, code generation, or research summarization is directly substituting for the entry-level labor that used to staff those functions. The companies building and deploying these systems are making a choice about which tasks to automate and which to augment, and the Stanford data shows the two paths diverge sharply in their labor outcomes.
Brynjolfsson frames this as a design problem rather than a prediction problem. “Outcomes depend on choices — by companies, policymakers, and workers,” he said. For AI builders, that means the decision to ship a tool that replaces a junior analyst’s workflow versus one that amplifies it is not a neutral technical choice. It is a labor-market intervention with measurable consequences for a specific demographic.
The most exposed occupations, writers at 57% and programmers at 55%, are the core constituencies of the AI industry itself. The people building and marketing these tools are, by the Tufts index’s own logic, among the most likely to be displaced by them. That is the paradox Chakravorti identified, and it is not abstract. The Stanford data on 22-to-25-year-olds is the first concrete evidence that the paradox is already resolving in real payroll data, not just in model evaluations.
The full wave has not arrived. Most workers still barely use these tools, and the effects measured so far are the leading edge. But the direction of travel is clear from the data already in hand: the next automation wave targets the youngest, most cognitively exposed workers first, and the industry building that wave is its own first customer.