Addy Osmani argues that the default way developers use AI coding tools optimizes for closing tasks rather than building understanding. Drawing on three recent studies — an Anthropic randomized trial, MIT's 'Your Brain on ChatGPT' EEG research, and a CHI 2026 paper — he shows that passive AI use measurably degrades comprehension and critical thinking. The key finding: the tool doesn't determine the outcome, the posture does. Engineers who used AI to ask conceptual questions outperformed those who copy-pasted generated code. Osmani offers concrete workflow shifts: form a hypothesis before prompting, ask for explanations before code, use learning modes (Claude's Learning Mode, ChatGPT's Study Mode), treat AI output like a junior engineer's PR, and periodically re-derive code by hand. He frames ship velocity and skill growth as two separate metrics — only one of which your employer tracks — and warns that engineers who can only ship with AI are entering a labor market already repricing what expertise is worth.

8m read timeFrom addyosmani.com
Post cover image
Table of contents
The studies are converging on the same pointThe tools default to shipping, not teaching“If the AI can do it, why do I need to understand it?”The fix is in how you prompt, not whether you doTwo metrics, not one
16 Comments

Sort: