---
id: "lib-2026-08-04-does-genai-narrow-education-productivity-gaps"
type: "library"
title: "Does GenAI narrow education productivity gaps"
lang: "EN"
org: "arXiv"
source_type: "Paper"
date: "2026-08-04"
week: "2026-W32"
url: "https://nosov.ai/library/lib-2026-08-04-does-genai-narrow-education-productivity-gaps/"
original_url: "https://arxiv.org/pdf/2608.04198"
tier: "A"
tier_name: "Strong evidence"
tier_reason: "Preregistered randomized experiment with 1,174 adults by economists at Universidad de San Andres, Nottingham, Maryland/NBER and Torcuato Di Tella, with chat-log analysis and an unassisted follow-up test."
summary_basis: "full-text"
original_minutes: "60"
tags: ["productivity", "rct"]
frames: ["solow-paradox"]
---

# Does GenAI narrow education productivity gaps

Tier A (Strong evidence): Preregistered randomized experiment with 1,174 adults by economists at Universidad de San Andres, Nottingham, Maryland/NBER and Torcuato Di Tella, with chat-log analysis and an unassisted follow-up test.

Summary based on the full text.

## In short

Guillermo Cruces, Diego Fernández Meijide, Sebastian Galiani, Ramiro Gálvez and María Lombardi (May 2026) ran a preregistered online experiment in Argentina to test whether generative AI narrows the productivity gap between people with more and less formal education. Participants solved a workplace-style business problem with or without a GPT-4.1 assistant, then answered questions without AI. AI raised performance for everyone, more so for lower-education participants, closing about three quarters of the gap. A gap remained, and part of it re-emerged once AI was removed.

## Key numbers

- 0.548 SD to 0.139 SD: education performance gap without vs with AI access, about 75% closed
- 1.242 SD vs 0.834 SD: AI effect on task score for lower- vs higher-education participants
- 0.171 SD: unassisted follow-up gain for lower-education participants after AI use
- 1,174: participants aged 25-45 in Argentina

## What it reports

- Without AI, higher-education participants score 0.548 SD above lower-education participants; with AI, the gap is 0.139 SD, closing 75% of it.
- AI raises the task score by 1.242 SD for lower-education and 0.834 SD for higher-education participants.
- Treated participants' scores remain dispersed, so the result is not everyone converging on a perfect score.
- In the follow-up without AI, treated participants do no worse than controls; lower-education participants gain about 0.171 SD, higher-education participants a small, non-significant amount.
- Lower-education treated participants still trail higher-education treated participants by 0.200 SD in the follow-up.
- Intensive AI use produces strong task scores even with low personal effort, but follow-up scores are higher only when intensive AI use is combined with high engagement in the task.
- Chat logs show both groups get substantial help; higher-education users write more detailed prompts and structure the workflow better.

## How they got it

1,174 Argentine adults aged 25-45, split by a preregistered education threshold (high school or under half of a tertiary programme vs more than half). The task: reply to an email from a boss by analysing several information sources, diagnosing a problem and proposing a solution, with incentive payments. The treatment group had a GPT-4.1 assistant embedded in the interface. All participants then completed an incentivised follow-up without AI. The trial is registered as AEA RCT 0016607.

## Limits

- Online, single-session task outside firms; the design trades organisational realism for variation in education.
- The follow-up is immediate, so it does not show skill retention over weeks.
- One country and one task type.

## Why it matters for companies

The result supports opening roles that previously required degrees, provided AI assistance is part of the job design. It also shows the limit: people with stronger baseline skills still use the tool better, and gains shrink when the tool is removed. Training on how to brief and structure work with AI, and keeping people engaged rather than handing over whole tasks, appear to decide whether gains carry over.

Original: https://arxiv.org/pdf/2608.04198