Most of us use AI the same way: describe the problem, read the answer, move on. That is useful when the job is getting something finished. Learning needs a little more friction.
That friction matters because a new preprint tested AI teachers across 1,500 simulated problems in math, code debugging, and brain teasers. The models tended to step in early and often. They also gave away more of the solution than people did in a smaller human-teacher comparison. The help improved immediate accuracy, but it did not reliably help the simulated students solve related problems later.
And that study cannot tell us how every person learns with every AI tool. A separate peer-reviewed survey of 319 knowledge workers found another caution. Higher confidence in AI was associated with less self-reported critical-thinking activity. That was an association, rather than proof that AI caused anyone to lose a skill.
So the practical distinction is useful. Sometimes you need the answer. Sometimes you need the attempt.
The next-clue routine
When you want to understand the work well enough to do it again, try this:
- Make your own start. Write what you know, what you tried, and where it stopped making sense.
- Show the attempt. Ask: “Give me one clue that helps me continue without finishing the problem for me.”
- Try again. Use the clue before opening another prompt. Keep the new attempt, including the part that failed.
- Choose the next kind of help. Ask for another clue if practice still matters. Ask for the full solution when completion matters more, then have the AI explain the reasoning and give you a fresh example to try alone.
This works for homework, a spreadsheet formula, a coding bug, a new process at work, or the sourdough starter that has chosen violence.
And before you ask AI for help, decide which job you are giving it. Speed is a valid job. So is helping you build enough skill to handle the next problem yourself.