Can AI narrow the education performance gap?

This paper reports that generative AI narrows the performance gap between higher- and lower-education people in a randomized online experiment run outside firms. 1,174 adults aged 25–45 completed an incentivized, workplace-style business problem-solving task with or without an AI assistant. "AI increases performance for all participants, with substantially larger gains for lower-education individuals." Without AI, higher-education participants outperformed lower-education participants by 0.548 standard deviations; with AI the gap fell to 0.139, "closing about three-quarters of the initial gap."

The design answers a selection problem the authors name. Existing evidence studies AI "within firms and occupations, where organizational selection compresses educational heterogeneity," so it cannot say whether AI narrows gaps across people with different levels of education. Recruiting outside firms restores that variation. The authors also separate two readings of the gains, productive use of the tool versus "mere delegation, fading once AI is unavailable," by adding a non-AI-assisted follow-up module. There, treated participants do not perform worse than controls and lower-education participants "retain part of their gain, although a sizable education gap remains," so the results "do not support a pure temporary-delegation interpretation." The third finding qualifies this: intensive AI assistance predicts strong main-task performance, but follow-up performance is substantially higher when assistance is combined with "sustained task engagement." Carry-over, on this account, depends on how the tool is used and not only on whether it is used. The paper frames all of this as evidence on an open dispute between Autor's "skill-democratizing" view of AI and Acemoglu's more cautious one.

Against the nearest notes, this is a partial contrast. Does AI assistance help workers learn lasting skills? reports gains that vanished on later unassisted work, while this paper finds no penalty and some retention for lower-education participants. The excerpt does not say how closely the follow-up module resembles the main task, so the two cannot be reconciled from here. The engagement finding echoes Does AI assistance actually harm the way developers learn?, where engaged use preserved learning, though the populations and tasks differ. On inequality, this is one experimental point on the narrowing branch of the two-sided picture in Does generative AI inevitably worsen or reduce inequality?. It shows narrowing happening in one setting without showing that narrowing is the general outcome. When does AI actually boost worker productivity? bounds the reading further, since this task is workplace-style problem solving and not the learning of a new skill.

The excerpt does not establish several things. It says there are "four main findings" but states only three, so the fourth is missing here. It gives no detail on the AI tool, the task content, how education groups were defined, the follow-up effect sizes, or how much of the gain was retained. It says intensive assistance "predicts" outcomes and does not say engagement was randomized, so the engagement result reads as an association. Nothing here speaks to real employment or wages, and durability is observed only through one follow-up module. At the strength supported, education-based performance gaps were not fixed under AI assistance in this setting, but the narrowing applies to AI-assisted performance, and in the unassisted follow-up a sizable gap remained.