The End of
Fake Learning
When finishing the work
stopped proving you learned it
The grade sees the result.
Learning happens in the process.
the homework
What the grade can see.
Where the learning happens. The grade can’t see it.
Same result. Same grade. No process, no learning.
Real learning is what you can still do:
without the help, later, on a new problem.
No tutorial.
not right after.
A new constraint.
Doing well with the help, right after, on the same problem is performance.
Fake learning is mistaking “I made it” for “I learned it.”
Same physics course.
Two ways to teach it.
A clear, polished lecture from a highly rated instructor.
Work the problems in small groups. Get stuck. Then the explanation.
01Which group learned more?
02Which group felt it learned more?
Your feeling of learning
is often backwards.
Less
More · +0.46 SDLess · −0.56 SD
Smooth made people think they had learned. Struggle made them think they hadn’t.
Often it’s the reverse.
The more you get wrong,
the more you learn.
An error teaches only when something tells you what went wrong. So smooth learning, with nothing to get wrong, often teaches nothing.
AI is most dangerous
when it’s right.
Each is smoother than the last, and leaves you less chance to be wrong.
When AI is right, you think you learned it too.
The practice got better.
The learning got worse.
Practice scores
with GPT-4 on request
Exam scores
once GPT-4 was taken away
Ted Chiang: it’s like bringing a forklift into the weight room.
The weights moved. The muscle didn’t grow.
Both fool your
sense of progress.
Visible activity standing in for value that’s hard to measure.
Visible results standing in for growth that’s hard to measure.
I relayed strategy down and results up. Good at it. Comfortable.
Then: what progress am I actually making?
School trains you to satisfy the grader.
Fake learning trains you for fake work.
Statsig, later acquired by OpenAI
Founder, Superlinear Academy
A new tool feels like progress.
It expires in months.
A new tool. An automation.
A finished tutorial.
Visible: screenshot it, post it, put it on your résumé.
Habits, methods,
ways of thinking.
Split a fuzzy task into checkable steps. When it fails, find the cause instead of switching tools.
The tool feels most like progress, and expires fastest.
The method feels like nothing, and stays with you. Backwards again.
A demo only has to be right at the demo.
A product has to stay right.
Clear requirements.
Simple data.
One user: you.
Inputs you never imagined.
Data that breaks.
And you’re not in the room.
Every step from prototype to product is your idea hitting reality.
Many right decisions don’t exist until you’ve seen the wrong version.
Where real learning lives now: from a tool to a method, from a prototype to a product.
Same AI. Different design.
Different learning.
Answers on request
Hints, not answers
The difference: a chance to think first, and to be wrong.
Bastani et al. · PNAS, 2025 · tutor version used teacher-supplied solutions and common mistakes ↗Where it counts,
choose the hard path.
Setting up environments, tokens, ports.
Burns willpower. Grows nothing.
Is this worth doing? How should it be split?
Is the result actually good?
Write your plan and where it will break. Then let AI do it.
A day later, redo it without AI, or explain it to someone.
New data, a new constraint, a different user.
Slower and harder? That isn’t falling behind. That’s learning happening.
When anyone can make beautiful work,
the work stops proving anything.
Fake learning booms first. Then it stops paying.
What still proves you: what you can do once the help is gone.
What ends is
a misunderstanding.
Mistaking “I made it” for “I learned it.”
Mistaking smooth for progress.
Something you learned with AI this month:
close the AI, change the problem. Can you still do it?