Using AI to complete coding tasks while learning a new skill comes at a cost: developers who relied on AI assistance failed to develop real understanding of what they were building. In randomized experiments where participants learned a new Python library, those using AI scored 17% lower on comprehension tests than those coding independently, nearly two grade points, with no significant time savings to show for it. While AI users who fully delegated code generation gained modest speed benefits, they never acquired the conceptual understanding, code reading, and debugging skills needed to evaluate their own work.
The researchers identified six distinct usage patterns. Three patterns preserved learning: asking only conceptual questions, requesting explanations alongside generated code, or generating code then following up to understand it. These required deliberate cognitive engagement. The other three, pure delegation, progressive over-reliance, and iterative debugging without understanding, produced quiz scores below 40%. The control group encountered three times as many errors during coding, forcing them to develop debugging skills the AI group never acquired.
The findings challenge the assumption that AI-enhanced productivity accelerates competence. When workers outsource problem-solving during skill acquisition, they complete tasks without developing the judgment needed to supervise AI in the future, particularly concerning for safety-critical domains.