The Quiz Test: How Geoffrey Litt Stops Cognitive Debt From Piling Up in AI Code

If you can't pass a quiz about the code your AI agent wrote, you don't understand it well enough to ship it. That is the rule Geoffrey Litt, a design engineer at Notion, described in a talk clipped here by @theaimemo. The clip pulls from Litt's presentation at the AI Engineer World's Fair 2026, where he argued that the real bottleneck in AI-assisted development is no longer whether the code works, but whether the humans on the team still understand it. The frames show his actual slides: a definition of "cognitive debt," a "Quiz me on the code" interface, and a diagram labeled "A quiz is a speed regulator."

The Failure Mode: You Stop Being Able to Participate

Litt opens by naming a pattern anyone who has vibe coded will recognize. Things go well, the agent produces working software, and then at some point you realize you have no idea what is going on. In his words, "I basically can't participate anymore." You have handed enough decisions to the agent that you can no longer reason about your own project.

The slide behind him in the frames calls this "cognitive debt" and quotes Margaret-Anne Storey: even if AI agents produce code that could be easy to understand, the humans involved may have simply lost the plot and may not understand what the program is supposed to do, how their intentions were implemented, or how to change it. That quote is real, and the slide cites Storey's blog directly.

The Fix: A Quiz at the Bottom of Every Explainer

Litt's countermeasure is concrete. He uses AI-generated code explainers, and at the bottom of each one there is a quiz. Five questions, medium difficulty. His rule is that he does not send code to teammates for review unless he can pass the quiz about what his agents wrote.

The frames show the mechanism in action. One slide reads "Quiz me on the code" over a document with numbered multiple-choice questions about a specific code change, such as which change caused a text box to expand and why. The quiz is not decoration. It is a gate that sits between the agent finishing and Litt asking another human to spend time on the result.

Why This Works: It Catches the Gaps You Didn't Notice

Litt admits the idea sounds silly, then says you should try it anyway. His reason is that it keeps catching him. It is "shocking the number of times this has caught me and made me realize I didn't understand." The value is not the quiz score. It is the moment of discovering a gap you would otherwise have shipped past.

This is the difference between passive review and active recall. Reading AI-generated code and nodding along produces a feeling of understanding. Answering a question about why one specific line behaves the way it does either confirms the understanding or exposes that it was never there.

The Bigger Idea: Speed of Understanding, Not Just Speed of Correctness

Litt frames the quiz as a "speed regulator." The slide in the closing frames shows a loop diagram: "Do I actually understand?" cycling around, with the quiz acting as the check. His point is that every incentive in AI tooling pushes toward going faster. Speed up, speed up, speed up.

His question reframes the goal. The team should not only move at the speed of correctness, meaning code that passes and works, but also at the speed of understanding. The quiz is the throttle that keeps those two speeds from diverging. When understanding falls behind, the quiz fails, and you slow down before the cognitive debt compounds into a project you can no longer change.

Key Takeaways

Cognitive debt is the gap between how fast AI ships code and how fast the team builds a real mental model of it, and it can leave you unable to change your own project.
Litt's rule: do not pass AI-written code to human reviewers until you can pass a five-question, medium-difficulty quiz about what the agent wrote.
The quiz turns passive code review into active recall, which surfaces the understanding gaps that reading alone hides.
Working code is not the same as understood code, so correctness is a weaker bar than comprehension.
Treat the quiz as a speed regulator: it lets you move fast while forcing understanding to keep pace with output.
The concept traces to Margaret-Anne Storey's work on cognitive debt as a bigger long-term threat than technical debt in AI-driven development.