The article makes a case that letting AI write code without active learning degrades the very skills engineers get paid for. Three recent studies — Anthropic, MIT, and CHI 2026 — all converge on the same conclusion: default AI usage without deliberate learning intent quietly erodes cognitive capability.

Anthropic's 2026 trial compared engineers learning Python with and without AI. Both groups finished tasks at equal speed, but the AI group scored 50% on a follow-up comprehension quiz versus 67% for the manual group. Inside the AI group, those who asked conceptual questions scored above 65% while copy-pasters scored under 40%. MIT's EEG study on essay writing showed brain connectivity scaling down with each layer of external support, with 83% of LLM users unable to quote their own output.

The fix isn't avoiding AI but changing how it's used. Practical shifts include forming hypotheses before prompting, requesting explanations before code, and turning on learning features like Claude's Learning Mode. Treat AI output like a junior engineer's PR — read it, critique it, push back. Ship speed and actual learning are two separate metrics; the defaults optimize for only one of them.