Research check · 18 October 2026
AI narrowed a task-performance gap. It did not prove a better job offer.
In an Argentine randomized experiment, adults with less formal education gained more from a GPT-4.1 assistant on one business problem-solving task. The measured education gap narrowed by about three-quarters while the tool was available. No employer hired anyone in this experiment.
What changed
Researchers randomly gave 1,174 adults aged 25–45 access to an AI assistant or no assistant. Participants analyzed a hypothetical business problem and drafted a response. Without AI, the higher-education group scored 0.548 standard deviations above the lower-education group. With AI, the gap was 0.139 standard deviations. Both groups improved; the lower-education group gained more.
What the test can and cannot show
Random assignment supports a causal claim about access to the assistant on this task. An immediate unassisted follow-up did not show worse performance for the assisted group; that is not proof of lasting learning. The task happened online, outside a workplace. The researchers did not measure job offers, pay, staffing or whether an employer would share access to the tool.
A narrower performance gap could change who can complete some work. It could also change how employers price or allocate that work. This experiment does not tell us which happens. A study linking tool access and use to actual hiring and wages would answer a different, more consequential question.
Source and version
We inspected the authors' full paper dated May 2026. The NBER catalogue lists an October revision, but its full revised text was not available to this check. The estimate here comes from the inspected paper, not an assumed change in that revision.