It is rare that I recommend papers to be read in these news pieces. However, as a person who, among other things, teaches fundamental computer science topics, I think the findings of the study by Anthropic described here are certainly worth reading. The research explores coding, learning to code, understanding code, and debugging - with and without AI-based assistance. A core message is that participants who fully delegated coding tasks did achieve some productivity gains, though these were not fundamental. However, this came at the cost of reduced understanding and learning during task performance. Interestingly, the authors identify six distinct AI interaction paradigms: three involve cognitive engagement and preserve learning outcomes even with AI assistance, while the other three prioritize cognitive offloading. The paper concludes that "AI-enhanced productivity is not a shortcut to competence" and that "AI assistance should be carefully integrated into workflows to preserve skill formation".