Education

AI narrowed a task-performance gap. Learning still required effort.

By Saint Zero Day · 2026-10-05 · 2 min read

A randomized experiment with 1,174 adults offers a more specific account of who benefits from assistance.

A working paper studies 1,174 adults aged 25–45 completing an incentivized workplace-style problem-solving task. Participants were randomly assigned access to an AI assistant, followed by a module without assistance. The paper appeared at NBER earlier in 2026 and on arXiv in August. [1] [2]

The authors report larger assisted-task gains for participants with less education. The measured gap between education groups fell from 0.548 standard deviations without AI to 0.139 with it, approximately a three-quarter reduction in that experiment. [1] [2]

Once the assistant was removed, a substantial gap returned, although the lower-education group retained part of its gain. The authors also associate stronger follow-up performance with sustained personal effort alongside intensive tool use. A randomized assignment to AI does not make every observed chat behavior a separately proven causal mechanism. [1] [2]

Why it matters

The useful interpretation is narrower than saying that AI eliminates differences in education. Assistance can make a specific task more accessible. Whether that becomes lasting skill, better employment or higher income requires different evidence. This is a working paper about an online task, not a classroom or workplace rollout.

For teachers and people designing workplace training, the follow-up is the interesting part. An excellent answer produced with assistance and an ability to work independently are different achievements. A training exercise should make space to see both.

Our editorial takeaway is to treat access and practice as complementary questions. Who can use the tool well? What do they understand afterward? The study offers a reason to pursue wider access while resisting the easy claim that handing someone a chatbot completes the educational work.