AI narrows the performance gap between education groups
Guillermo Cruces, Diego Fernández Meijide, Sebastian Galiani, Ramiro H. Gálvez et al. · Cruces et al. · arXiv / NBER / IDB · 4 Aug 2026
Source finding
Adults used supplied information to diagnose a business case and propose a solution by email. The assessed performance gap between higher- and lower-education groups was 0.548 standard deviations without AI and 0.139 with AI.
Our interpretation
AI can narrow differences in the current work result. This establishes neither equal independent ability nor a corresponding business value.
Method and evidence
Preregistered randomized online experiment with 1,174 adults aged 25–45 in Argentina, September–November 2025. Same self-contained business case, with or without an AI assistant. AI-assisted grading checked against a human-graded subsample and alternative scoring procedures. Strong evidence for this short-term task comparison; working paper, not peer-reviewed.
Limitations and interests
Outside firms, not real workflows. Education groups defined by progress through postsecondary education; education is not a complete ability measure. The unassisted follow-up was immediate, not a test of long-term learning. Lower-education participants retained part of their gain; a sizable education gap remained. Support from Sur Futuro, IDRC and the Inter-American Development Bank; no workplace productivity test.
What businesses can do
Assess AI-assisted results and independent ability separately. Ask employees to solve a similar case later without AI before changing training or task allocation.
Where to start
Section 2 for the case and education groups, Table 1 for performance gaps, section 5 for unassisted follow-up. Original report, October 2026