Don't Outsource Learning

Addy Osmani's May 2026 essay names the failure mode that emerges when an engineer settles into the default AI coding loop: paste spec or error, accept the fix, ship, repeat. The symptom vanishes each time, but the messy struggle between problem and solution stops happening at all. Across thousands of these micro-interactions, what you can build without AI watching gets a little weaker every week. None of these moments feel like a problem on the day they happen.

Osmani is not anti-AI — he uses these tools daily and has shipped more in the last year than the five before it. The argument is about posture: the default UX is tuned for closing tasks, which is a different goal than staying sharp enough to steer the tools over a multi-decade career.

The three studies that converge

Three independent pieces of research published in the year leading up to the essay land in roughly the same place:

Anthropic (early 2026) ran a randomised trial where engineers learned a new Python library, half with AI assistance and half without. Both groups finished at the same speed. The AI group bombed the follow-up: 50% on the comprehension quiz vs 67% for the manual group, with the gap widening on debugging. The interesting cut was inside the AI group: engineers who asked the model conceptual questions scored above 65%, those who copy-pasted the generated code scored under 40%. The tool didn't determine the outcome — the posture did.
MIT's "Your Brain on ChatGPT" compared essay writing across LLM, search-engine, and brain-only groups using EEG. Brain connectivity scaled down with every layer of external support. The LLM group showed the weakest coupling. After writing, 83% of LLM users couldn't quote a single line of what they had just produced. The researchers named the pattern cognitive-debt: saving mental effort today, paying for it in critical thinking tomorrow.
CHI 2026 added the ordering finding. When people had LLM access at the start of a task, the LLM framed the entire problem. Even when the human did the rest of the work themselves, that initial anchoring produced measurably worse decisions. Total AI use mattered less than where in the workflow it appeared.

Different methodologies, same conclusion. Osmani's framing: using AI without an active intent to learn quietly degrades the skill you're being paid for.

Why the defaults push the wrong way

If you fire up a coding agent and stick to the defaults, the loop is optimised for one metric: getting the task done. The model writes the code, you accept, repeat. At no point does the tool pause and ask "what do you think the problem is?" or "try writing the first five lines yourself."

Osmani's read is that this isn't a conspiracy — it's UX gravity. Product teams are rewarded for merged changes and shorter cycle times, not for making engineers sharper. We all want fewer keystrokes, so the tools have sanded friction away. The trouble is that friction was where the learning lived.

A few vendors have pushed back: Anthropic shipped Learning Mode for Claude (Socratic questioning, asks you to write code before continuing), OpenAI shipped Study Mode, Google shipped Guided Learning. Almost nobody uses them for real production work. We've filed them under "for students" — Osmani calls this a mistake. The same feature that helps a sophomore learn React works for a senior engineer learning Rust. You just have to be willing to feel like a beginner again.

When delegation breaks down

Osmani concedes the fair question: "if the AI can do it, why do I need to understand it?" For boilerplate, glue code, throwaway CI scripts — sure, delegate. The opportunity cost of memorising YAML is too high. But pure delegation fails in five specific places:

When something breaks. AI-generated code crashes the same way human code does. "The agent wrote it" doesn't debug anything.
When it's confidently wrong. The only defence against a plausible-looking incorrect answer is enough expertise to spot it.
When the foundation changes. Frameworks update, security reviews flag structural issues — you can't re-prompt your way out of a migration.
When you leave the median. LLMs are brilliant at problems solved a million times on GitHub. The further from median, the worse they get. The undocumented hard problems still require deep understanding.
When the market adjusts. The 20% drop in junior developer employment since 2022 is the labour pool already re-pricing what expertise is worth.

The posture-shift prescription

The same tools can produce sharper engineers. The difference is in what you ask:

Form a hypothesis before you ask. Two or three sentences on what you think the problem is. Use the model's answer to test your theory, not to replace it.
Ask for the explanation before the code. In unfamiliar territory, the first prompt is "explain how this works, what the alternatives are, what the tradeoffs are." Code comes after.
Turn on Learning Mode when you're out of your depth. Yes, it's slower. That's the point.
Treat AI output like a PR from a junior engineer. Read it, critique it, push back on it.
Re-derive things by hand once in a while. Take a piece the model wrote, recreate it from scratch. It's the calibration check that tells you how much you've quietly lost.
Ask the model to teach you what it just did. One extra prompt changes what you take away from the session.

None are dramatic. They're small posture shifts inside tools you're already using.

Two metrics

Osmani's closer: end each coding session with "did I learn anything today, or did I just close tickets?" Sometimes the honest answer is the second one, and that's fine. If it becomes the answer for months in a row, cognitive-debt is accumulating in the background.

Ship and learn are two separate metrics. Your manager and your customers only ever ask about the first. The second is on you.

Cross-references

cognitive-debt — the concept the MIT study names; the structural argument for the prescription
skill-atrophy-supervision-paradox — Lars Faye's heavier framing of the same dynamic at the senior-engineer / agent-supervisor level
agentic-coding-fatigue — 0xsid's decision-fatigue framing; the missing piece is that the fatigue compounds while comprehension declines
agentic-coding-is-a-trap — supervision requires the skills agent use erodes; Osmani's prescription is the practical "stay sharp" answer to that trap
building-syntaqlite-ai — the same posture argument from the project side: 8 years of wanting, 3 months of disciplined design with AI
no-silver-bullet-llms — Bennett's argument that DORA/CircleCI data already shows the productivity claim fading; same direction, different evidence
claude-emotion-concepts — interpretability work on the model side; the human-side companion piece is to keep your own concepts intact
ai-assisted-workflow — Barbero's seven-step planning workflow that puts thinking before code; the procedural complement to Osmani's "form a hypothesis first"