Fed paper flags AI job losses

A Federal Reserve paper says large language models have already erased about 500,000 U.S. coding jobs that likely would have existed without them. (federalreserve.gov) The paper, published March 20, 2026 by Federal Reserve Board economists Leland D. Crane and Paul E. Soto, tracks programming-heavy occupations using Occupational Information Network data linked to the Current Population Survey. It finds coder employment “decelerated sharply” after ChatGPT’s release in November 2022. (federalreserve.gov) The authors estimate the gap by comparing post-2022 coder hiring with earlier trends and adjusting for broader industry slowdowns. Using 5.735 million coder jobs as the base, they write that roughly 500,000 additional jobs would have existed without large-scale large language model use. (federalreserve.gov) Large language models are prediction engines trained on vast text and code archives; in software work, they autocomplete functions, tests, and documentation much like a more capable version of code suggestions. The Fed paper focuses on coding because outside usage data shows programming is one of the most common uses of generative artificial intelligence. (federalreserve.gov) The paper is a working paper, not an official Federal Reserve policy statement, and the Board says its conclusions are preliminary and circulated for discussion. Even so, it adds federal research weight to a labor-market shift that has mostly been measured through company anecdotes and sector surveys. (federalreserve.gov) Other Federal Reserve research points in the same direction on who is most exposed. A Dallas Fed analysis published February 24, 2026 says artificial intelligence appears more likely to substitute for entry-level workers handling codified tasks and to complement experienced workers whose value comes from tacit knowledge gained on the job. (dallasfed.org) That split matches the career ladder in programming, where junior engineers often do debugging, testing, and boilerplate work that can be standardized, while senior engineers are paid more for system design, judgment, and coordinating teams. Dallas Fed researchers also cited Stanford work finding employment declines in artificial-intelligence-exposed sectors were concentrated among workers under 25. (dallasfed.org) The broader picture is less one-sided than the coder numbers alone suggest. An Atlanta Fed summary of a survey of more than 700 corporate executives, published March 25, 2026, found companies reporting productivity gains from artificial intelligence and expecting workforce reshuffling, but “little evidence” of near-term economy-wide job cuts so far. (atlantafed.org) Federal labor projections still show growth in several computer occupations. The Bureau of Labor Statistics said on March 11, 2025 that software developer employment is projected to rise 17.9 percent from 2023 to 2033, partly because companies still need people to build artificial-intelligence systems and maintain more complex data infrastructure. (bls.gov) The immediate signal from the Fed paper is narrower: coding looks like the first major office job where large language models are visible in aggregate labor data. The next question is whether the tools keep replacing junior work faster than firms create new senior roles around them. (federalreserve.gov)