The 17% Tax
Anthropic's own RCT found developers using Claude scored 17% lower on comprehension. But one group, those who asked instead of delegated, performed just as well.
Anthropic studied its own product and published the uncomfortable result.
In a randomised controlled trial, developers using Claude to complete coding tasks scored 17% lower on comprehension tests. No significant speed gain. The velocity was an illusion.
But the study had a plot twist. Developers using “Conceptual Inquiry” mode (asking probing questions, challenging assumptions, requesting explanations) performed as well as unassisted developers, while completing tasks faster.
The difference was not the tool. It was the posture. Delegation atrophies skill. Inquiry preserves it.
Every AI interaction is a choice between these two modes. Most people are choosing wrong.