Grades and prior knowledge barely predicted who gained from AI. Skill in handling the model did.
A study by Anand and Idan, released as a preprint in 2026, had participants teach themselves a technical subject — half with traditional resources, half with an AI model. On average, AI helped. But the gains were uneven. Top grades and prior knowledge did little to predict who benefited. What did: the ability to ask the model the right questions, sort its answers, and check them. The authors call this AI Interaction Competence. People strong in it gained a lot. People weak in it gained little — some did worse. One simple aid, concept maps to structure the work, narrowed the gap.¹
The paper has not yet been peer-reviewed, so treat the exact numbers with caution. The direction matches what field experiments on AI at work keep finding.
Talent Takeaway: Before you buy more licenses, spend an afternoon teaching people to question the tool: ask, sort, check. A shared template is the cheapest way to keep the gap from widening.
¹ Anand, B., & Idan, L. (2026). Generative AI and the productivity divide: Human–AI complementarities in education and knowledge work [Preprint]. arXiv. https://doi.org/10.48550/arXiv.2605.18143
Further reading: Brynjolfsson, E., Li, D., & Raymond, L. (2023). Generative AI at work. https://doi.org/10.3386/w31161
