February 3, 2026
Anthropic Just Studied How AI Affects Learning to Code
AI can help you ship faster. Anthropic's study asks what happens to the learning underneath.
There is an uncomfortable difference between producing code and understanding it. AI makes the first one faster. It can make the second one optional.
A study from Anthropic gives that problem a useful shape. The sample was small and the test was narrow, so it does not settle the future of programming. It does show how quickly learning can disappear when output becomes the only measure of progress.
The study
Anthropic recruited 52 mostly junior developers and asked them to work with Trio, a Python library none of them had used before. One group could use an AI assistant. The other coded by hand.
| Group | Quiz Score | Time |
|---|---|---|
| No AI | 67% | baseline |
| With AI | 50% | about 2 minutes faster |
The AI-assisted group averaged 50% on the follow-up quiz. The hand-coding group averaged 67%. The difference was especially visible in debugging questions.
Two minutes is not much of a productivity gain. A 17-point gap in understanding is difficult to ignore. But this was an immediate quiz about an unfamiliar library, not a longitudinal study of professional developers. The researchers are explicit that they have not shown what happens to skills over months or years.
The patterns
The usage patterns explained more than the average score did.
Participants who delegated the implementation tended to finish quickly and understand less. Participants who asked conceptual questions, inspected generated code and followed up on explanations performed better. AI was present in both patterns. The difference was whether it replaced the learning loop or participated in it.
What this means for teams
I would not respond by banning AI. I would change what teams ask people to demonstrate.
A pull request is evidence that code exists. It is not evidence that its author can change it safely next week. Ask the developer to explain the failure mode. Let them reproduce a bug without delegating the diagnosis. Use reviews to expose reasoning, not just to approve a diff.
This matters most for juniors because routine implementation used to provide the repetitions that built intuition. If an agent performs those repetitions, teams need another way to create them. Otherwise we get faster delivery today and fewer people able to debug it tomorrow.
The study does not prove that AI makes developers worse. It shows that using AI without protecting comprehension can make output a poor proxy for skill. That is enough to change how I review work.