Two months ago I wrote something down in my second brain: "friction is the curriculum."
Not as a slogan. As a conclusion. Two essays had landed on Hacker News in the same week — one about hobby programmers who refuse LLMs because "the process of mastering the difficult field IS the product," and one called "Taste Is All That's Left." The argument: the production wall wasn't just an obstacle, it was the curriculum. Every wall you climb teaches you which walls are worth climbing. Remove the wall and you remove the apprenticeship. Taste isn't absorbed by reading good work — you build it "the slow, stupid, humiliating way": you make something bad, it fails in front of you, and some part of you files the failure away.
I wrote it down because I believed it. Then I mostly stopped thinking about it, because I had a course platform to build and a cohort to teach.
This month, the experiment arrived.
Philip Oreopoulos and Nina Low ran the first large randomized trial of an AI tutor — Khanmigo, in 18 middle schools in Hamilton County, Tennessee, across two school years and 6,902 student-term observations. Read past the headline and the result is brutal: the AI tutor added nothing. Students assigned Khanmigo gained 1.26 percentile ranks per term in math — a real, statistically significant gain, but statistically indistinguishable from what plain Khan Academy practice without AI already produces. The AI layer, by the authors' own measure, contributed zero.
The reason isn't that the model was bad. It's in the usage logs.
Ninety-six percent of students messaged the tutor at least once. But the median student used it on a third of the days they practiced, in 14% of their exercise sessions — and in only 17% of the sessions where they actually got a problem wrong.
Seventeen percent. The mistake is the entire moment a tutor exists for — the confusion, the "wait, what did I do wrong," the exact second a human tutor would lean in. Four times out of five, the student didn't open the tool.
And what they did send was worse. Of every message, 39.4% were bare answers, 24.2% were clicks on a suggested prompt, and only 14.5% contained an actual math question or a step of reasoning. They wanted the answer. When Khanmigo refused — it was configured to coach, to ask questions back, to withhold — students largely stopped using it.
Sal Khan, whose company built the thing, called Khanmigo a "non-event" for most students. The CEO.
Here's the part that got me. The fix the researchers found, in a companion study, is the thing I'd been arguing with a different name: force the engagement. When the AI popped up uninvited and gated progress behind a demonstrated attempt, students got slower and more accurate — and still remembered it a week later. Slower. That's the friction doing the teaching. The struggle wasn't the bug. It was the product.
And here's the part that stung.
My course platform — the one I teach a live cohort on — has a sprint sitting in the backlog called "AI Features." Two items: an OpenRouter Q&A for students, and AI submission summaries. Status: not started. The first item, the Q&A, is an opt-in "ask the AI" tab students click when they want help.
Which is, word for word, the design the study just measured and found to be furniture. The 2023 Khanmigo was a separate tab students had to click — the exact version the researchers found students ignored. Khan Academy has since rebuilt it so the tutor activates automatically after a wrong answer. The experiment proved my principle right one sprint before I would have shipped the feature that violates it.
I wrote "the wall is the teacher." Then I almost built a door with no wall.
So here's the rule I'm writing into the sprint brief before a single line of code gets written: no opt-in AI surface. No "ask the AI" button a student has to discover. The coach appears at the point of struggle — after a wrong answer, a failed check, a stalled submission — not on demand. It withholds the answer by structure, not by a system prompt a student can defeat by asking twice. And the submission summaries are for me, the instructor, to triage forty students in a weekend — never handed back to the student to replace the thinking I'm trying to teach.
My day job got the same lesson in dollars this year. On-target earnings rose six to nine percent, and more than three-quarters of sellers missed quota. The broken part wasn't the plan — it was the quota, set on a blanket rule instead of the conversion math already sitting in the CRM. Every number in the chain should be derivable, not asserted.
Same sentence, two domains: the input is never the lever. Access to an AI tutor isn't learning. A bigger target comp isn't attainment. You can't make the struggle optional and expect it to happen — you have to build it in, and you have to derive the number instead of asserting it.
I teach the first thing, and I run the team that has to live the second. Which makes the uncomfortable part hard to avoid: the person on this side of the screen with the most articulated philosophy about engagement was one sprint away from shipping an opt-in tab.
The experiment didn't tell me anything I didn't already know. It just timed its arrival so I couldn't pretend otherwise.
And that's the part I can't shake. If I'd shipped that tab, nobody would have noticed it wasn't working. Students would have ignored it quietly, and I'd have called it "done." That's what furniture does.