AI Shrinks the Training Gap in a Workplace Experiment

Generative AI can sharply, if temporarily, narrow performance gaps between people with different education levels. In a randomized experiment with 1,174 adults, the measured gap on a business problem-solving task shrank by roughly three quarters when participants could use an AI assistant. The follow-up is just as important: once the assistant disappeared, a significant share of the gap returned.

Researchers Guillermo Cruces, Diego Fernández Meijide, Sebastian Galiani, Ramiro Gálvez, and María Lombardi asked adults aged 25 to 45 to complete a work-like business problem task. Participants were financially incentivized to perform well. A randomly selected group could use a generative AI assistant; the control group did the same task without AI. Everyone then completed a second part with no assistant.

The study first appeared as an NBER working paper and was also posted on arXiv this week. That means the results are public and open for discussion, but don’t yet carry the weight of a finalized article that has passed a journal’s full peer-review process.

In the control group, participants with more education scored on average 0.548 standard deviations higher than those with less education. With AI, that difference fell to 0.139 standard deviations. In other words, about 75 percent of the initial gap closed. Both groups performed better with AI, but the relative gain was larger among the less-educated group.

Chat transcripts showed that less-educated participants extracted a lot of useful help from the assistant. More-educated participants, however, still used the system slightly more effectively on several fronts. AI pulled performance closer together, without eliminating differences in knowledge and approach.

The second, unsupported part adds nuance. Participants who had used AI first did not perform worse later than the control group. Among less-educated participants, some of the improvement stuck. Still, a clear education gap reappeared as soon as nobody had access to the assistant.

Personal effort mattered too. Intensive AI use correlated with better performance during the assisted task, regardless of how much time the participant invested. In the later task without AI, results were strongest when heavy AI use was paired with sustained personal effort. Simply copying an answer delivered less lasting benefit than actively working with the explanation and approach.

The most appealing takeaway is that generative AI can narrow knowledge gaps. An employee with less experience in a business analysis task can reach a usable result faster when an assistant structures information, suggests alternatives, and drafts a first version. That can help organizations tap talent more broadly and get new hires up to speed faster.

The riskiest conclusion would be that training and guidance are no longer needed. The study shows the underlying differences resurface once the tool is removed. Moreover, higher-educated participants used AI more effectively in some respects. An organization that merely provides chatbot access may mistake a temporary productivity boost for durable skill development.

Employers need a mix of access, practice, and oversight. Employees should learn what context a model needs, how to steer an answer, and how to verify numbers, sources, and assumptions. Our guide on AI literacy at work outlines the organizational agreements that support this. Our recent explainer on Article 4 of the EU AI Act and AI literacy also shows this is more than a productivity question.

Good implementation doesn’t start with “use AI from now on,” but with a well-defined work task. First, have employees describe how they would handle the task without help. Then run a structured AI exercise with relevant context and quality criteria. Next, discuss which parts of the AI’s answer were useful, what errors occurred, and which checks were necessary. Finish with a comparable task without the assistant to test what was actually learned.

That approach aligns with our AI at work practical guide, where employees and teams learn to integrate AI—reliably and verifiably—into their routines using checklists, fill-in exercises, and a two-week practice plan. If you mainly want to craft better prompts, start with our explainer on how to turn a mediocre ChatGPT answer into something useful.

One online experiment is not proof that AI erases three quarters of education gaps in every job. Participants completed a single type of business problem in a controlled setting. Results may differ for long-running projects, physical work, creative tasks, specialist advice, or tasks where mistakes carry high stakes. And a difference in standard deviations doesn’t directly translate into minutes saved, euros earned, or jobs created.

The value of the research lies less in a universal percentage and more in the mechanism it reveals. AI can bring people closer together during a task, especially when it compensates for a lack of experience or structure. But lasting development still requires understanding, practice, and personal effort.

For employers, the best investment isn’t just a license. It’s a learning process where employees show they can master the task both with and without AI.