Does AI assistance narrow performance gaps between education groups?
Core argument: In a randomized experiment, access to an AI assistant raised overall task scores by 1.242 standard deviations for low-education participants versus 0.834 SD for high-education participants, a gap of 0.408 SD significant at the 1% level.
The experimental task presented a realistic hypothetical business scenario - an email from the participant’s boss describing a problem [requiring] analysis of information sources, including text, a figure, and a table. Participants were asked to answer three questions to diagnose the problem and propose a solution. We assess performance in the task on a 0–10 scale along two dimensions, content and writing. Figure 2 presents the results on the impact of having access to the AI assistant on this outcome for each education group. The task without access to the AI assistant is substantially more challenging for low-education participants: their high-education counterparts perform 0.548 standard deviations (SD) higher in the overall score. AI substantially increases task performance for participants of both education groups - overall score rises by 1.242 SD for those in the low education group, whereas the impact of the assistant for high-education subjects is 0.834SD. The treatment effect of access to the AI assistant is 0.408 SD higher for the low-education group [than for] the high-education group, with both results statistically significant at the 1% level.




