An experiment: AI narrows the education gap
In February 2026 NBER issued a randomised experiment on 1,174 adults in Argentina: with a generative-AI assistant the gap in task performance between higher and lower education fell from 0.548 to 0.139 standard deviations.
Why it matters
A measurement, not a forecast: an AI assistant gives more to people with less education and closes about three quarters of the gap on the task. An editorial assessment: the task is general, so nothing follows about particular occupations.
What is said. An online experiment in Argentina, September-November 2025; 1,174 adults aged 25-45 do an incentivised business problem-solving task with or without a generative-AI assistant. AI raises everyone's result, more for the lower educated. Without AI, higher-educated participants outperform the others by 0.548 standard deviations; with AI, by 0.139. The authors stress that the experiment is outside firms, where selection compresses differences in education. What the record does not claim. The page and the PDF of the May revision were read; the earliest Internet Archive capture (14 March) has the same 1,174, 0.548, 0.139 and "three quarters", but a February text was not seen. The May version adds a follow-up task without AI and chat-log measures; they are not taken into the record. A working paper, not peer-reviewed.