Don't Outsource Your Understanding

I recently read Don’t Outsource Your Understanding by Han-chung Lee, and it put words to something I feel every time I ship with an AI agent. The argument is simple: you can outsource the work to a model, but the moment you also outsource the understanding, the debt comes due - and you won’t have enough of the problem in your head to pay it.

Offloading vs. surrender

Lee draws a line between two things that look similar but aren’t:

• Cognitive offloading - you let the tool do the grind, but you stay in the loop to read, check, and learn from the output. You still own the mental model.
• Cognitive surrender - you remove yourself entirely. The tool does the work and the thinking, and you just paste whatever comes out.

Offloading compounds into skill. Surrender compounds into a slop-field: layer after layer of plausible-looking, low-quality output stacked on top of each other until the whole thing slows to a crawl and nobody can say why.

When the bill arrives

The essay’s examples land because they’re ordinary, not exotic:

• A solo founder ships with Cursor, Lovable and Bolt for six months. The app works - real users, real revenue. Then a contractor opens the codebase and says “wtf is this.” No structure, no architecture, and a founder who surrendered the understanding can’t fix the slop.
• A lawyer files a brief with citations the model hallucinated.
• Engineers chain AI nodes into a pipeline and never test or evaluate any of it.

In every case the failure isn’t “the AI was bad.” It’s that the human stopped verifying.

My take, as an engineer

This maps directly onto how I try to work with agents day to day. The rule I keep coming back to: let the agent write it, but never let it be the last one to read it.

• Read the diff, not just the summary. Walk the code path the change actually takes.
• Re-derive the tricky part in your own words before you accept it. If you can’t explain why it works, it isn’t done.
• Treat verification as the real work - because it is. The reconstruction is where the understanding gets built.

Agents make me dramatically faster at the parts that were always tedious. What they can’t do is hold the mental model for me. Staying in the loop - reading, checking, rebuilding it in your own head - is the part worth protecting.

Source

This post is a summary and personal reflection. The core ideas - cognitive offloading vs. surrender, the “slop-field”, and the founder / lawyer / pipeline examples - all come from the original essay. Please read it in full:

• Han-chung Lee, “Don’t Outsource Your Understanding”, 1 May 2026. Available at: leehanchung.github.io/blogs/2026/05/01/dont-outsource-your-understanding