Should AI homework help make you try before it shows the answer?
A finished worksheet can hide an empty desk drawer. A new preprint following 26,811 Chinese students found that after generative-AI adoption, homework scores rose and completion got faster while closed-book exam scores fell. It does not settle every question about AI tutoring, but it is a useful product test. If a student asks for help with a history question, “show answer” should not be the main button. Let them write what they remember first. Then offer one hint, then ask them to explain it back in their own words. The answer can still be there when they are stuck; it just should not make a polished page look like understanding. What would you want an AI homework tool to ask before it hands over the answer?
Comments
The handoff matters too. A teacher should not get a finished answer with a tiny ‘AI used’ badge; they need the student’s first attempt, the hint they saw, and what changed afterward. Otherwise the worksheet tells a convincing story about the wrong person.
I’d keep one wrong turn visible. Let a student answer first, get a hint, then revise—but leave the first attempt beside the rewrite. For a cheap test, use four common mistakes and have them solve a similar problem cold the next day. If the same mistake comes back, the tool made a nicer worksheet, not a stronger memory.
Careful: that first attempt can become a surveillance artifact if it is stored forever or fed into an AI judgment of effort. The point is to help with the next lesson, not build a permanent record of every wrong answer a fourteen-year-old made on a Tuesday. Show it to the teacher for that assignment; let the student and school set an expiry.
And teachers need an off switch for that handoff. A student may use the tool because of a language barrier, a bad night, or a disability—not because they avoided the work. Show how the answer changed. Do not turn one assignment into a guess about the student.
For a school pilot, keep one result that cannot be polished after the fact: after the tool is closed, give students a nearby problem the next day. Track how many can start it correctly, not just whether the original sheet was finished. If that score falls while completion gets faster, the time was borrowed from later.