A student can now produce a tidy summary, a working chatbot or a polished slide deck in about ten minutes. That makes AI-assisted student work hard to grade. When the finished product costs almost nothing to make, it tells a teacher very little about what the student understands.
The fix is to grade the explanation. A student’s account of why they made each choice is hard to generate on their behalf, because it depends on what they actually did.
Students are already working this way. In a Pew Research Center survey taken in autumn 2024, 26% of US teens said they had used ChatGPT for schoolwork, double the share from a year earlier. Teachers who grade only the product are increasingly grading the tool.
Why grade the explanation instead of the output?
Explaining is one of the best-supported ways to learn. In their widely cited review of study techniques, Dunlosky and colleagues found that self-explanation helps across a range of subjects and ages. That covers learners explaining how new information connects to what they know, or why a step makes sense. So asking for an explanation is good assessment practice, and the act of explaining also teaches.
It also changes what counts as effort. A student who uses an AI tool and then explains accurately where it went wrong has done more thinking than one who wrote every word alone and cannot say why any of it is true.
5 explanation prompts for any AI-assisted assignment
These prompts turn AI-assisted student work into something you can assess. They work for a history essay, a science lab write-up or a first machine learning project. Pick two or three per assignment.
- What did you give the tool, and why that? Students describe the prompt, data or source material they started from, and what they left out.
- Where was the tool wrong or weak, and how did you notice? Every honest answer has at least one example. “Nowhere” usually means the student did not check.
- What did you change after the first attempt? This shows whether the student iterated or accepted the first output.
- What would break it? Ask for one input, question or situation where the result would fail.
- What would you do with another week? A specific answer shows the student understands the work’s limits.
A 0 to 2 rubric for scoring the answers
Keep scoring light. A three-point scale per prompt is enough, and it is quick to apply during a short conference or while reading a written reflection.
| Score | What the answer looks like |
|---|---|
| 0 | Generic. It could describe anyone’s project (“I made sure it was accurate”). |
| 1 | Specific to this project, but the student cannot back it up when asked. |
| 2 | Specific and checkable. The student points to a line, a source or a result and can show it. |
The gap between a 1 and a 2 is the one that tells you the most. Students often know the right words (“I checked for bias”). Fewer can show you where they did it.
One spot-check keeps the written answers honest. Collect them, then pick one prompt at random and ask it out loud for two minutes. If the spoken answer matches the written one, you can trust the written answers more for the rest of the term.
Where do first student AI projects usually stall?
Teachers running a first AI project with a class tend to hit the same two walls. Knowing them in advance makes them easier to plan around.
Messy data in the first week
Students expect the hard part to be the code or the model. Usually it is the data. The spreadsheet has blank rows, labels mean different things in different columns, or the “dataset” turns out to be fifty rows. Students lose momentum here faster than anywhere else, because nothing they do looks like progress.
Build a cleaning day into the plan and grade it. Asking “what did you have to fix in your data, and why?” turns the slowest week into something you can assess with the prompts above.
A question too big to answer
“Can AI predict climate change?” is a great interest and an impossible first project. Help students shrink the question until one dataset can answer it, such as “Can rainfall data from our county predict which months the creek floods?” Showing students two or three examples of beginner-sized AI projects can help them see what “small enough” looks like.
A good rule of thumb: if the student cannot describe what a correct answer would look like, the question is still too big.
How to start grading AI-assisted student work this week
You don’t need a new policy to try this. Take one assignment you already give, allow AI help, and add two explanation prompts and the 0 to 2 scale. Then compare what you learn about each student with what the finished product alone would have told you.
If the explanations show you something the product didn’t, such as which students checked the tool’s work and which only finished the task, you have your case for doing it again.
Bio: Maria is a programme manager at BetterMind Labs, which runs mentor-guided AI and machine learning project programmes for students in grades 6 to 12. She works on how students explain and defend the things they build with AI.



