The AI productivity paradox: the more it helps your job, the more it can replace you

Consider a junior copywriter who now drafts three blog posts an hour with AI help. She is faster, more productive, and, at the same time, near the top of a new ranking of 784 U.S. occupations by AI task overlap. The Tufts Digital Planet American AI Jobs Risk Index puts writers and authors at 57% task overlap with current AI systems, the highest score on the list, followed by computer programmers and web and digital interface designers at 55%, editors at 54%, and web developers at 46%. The mechanism is the same: the more cleanly a job decomposes into tasks an AI can already do, the more an AI can also do the job.

The index, released in March 2026 by Fletcher's Digital Planet initiative, ranks 784 occupations across 20 industry sectors, 530 metros, and 50 states by AI-driven job-loss vulnerability, drawing on AI capability benchmarks and labor data rather than a survey of human behavior. Under a median adoption scenario, the index projects 9.3 million U.S. jobs at risk, with a range of 2.7 million to 19.5 million depending on how fast AI capability diffuses. The midpoint income at risk works out to about $757 billion. That is not a forecast of layoffs; it is a measure of how much current AI capability overlaps with what workers get paid to do.

Bhaskar Chakravorti, dean of global business at Tufts' Fletcher School and one of the researchers behind the study, calls this the labor market paradox: the more AI helps your job, the more expendable you become. Tech hubs that benefit most from AI productivity gains also sit at the top of the exposure ranking. Software developers, management analysts, market research analysts, and marketing specialists each combine high salaries with large headcounts, which means the absolute income at risk is concentrated in exactly the roles that look most resilient in casual coverage. The same job that looks safest on a "future-proof your career" list is the job with the most substitutable task structure.

The first measurable shock in U.S. payroll data is already visible, and it is not landing where most forecasts expected. A Stanford Digital Economy Lab analysis of ADP payroll records covering millions of workers finds that the earliest labor-market effects are concentrated among workers in their twenties, a leading edge that mirrors the occupational ranking. Junior roles are typically narrower in scope and heavier in tasks AI can already execute, which the lab's analysis points to as the reason early-career cohorts are absorbing the first hit, though the published data does not yet spell out the full causal mechanism. The index puts 4.9 million workers across 33 "tipping point" occupations in the same category, and PR specialists, news analysts, reporters, journalists, and market research analysts each score between 35% and 37% task overlap.

A December 2025 executive order directed the Department of Justice to challenge state AI laws, and the IBGC press release accompanying the index notes that BEAD broadband funding is being deployed in states most exposed to AI job loss, which also happen to be the states most active on AI regulation. The geographic pattern matters because federal leverage is being applied to the same metros the index flags as highest risk.

The ranking measures task overlap and capability exposure, not realized job loss, and the productivity-augmentation rebuttal is real: workers who use AI as assistive review-and-edit support may move up the value stack rather than out of it. The first-draft substitution path looks different from the review-and-edit path, and the labor data cannot yet tell them apart. The Stanford payroll signal shows where the shock is landing, not whether each affected role ends up smaller, larger, or reorganized.

The next data point is whether the early-career signal broadens into mid-career cohorts over the next four quarters of ADP data. If it does, the productivity paradox stops being a paradox and becomes a forecast.