AI tutors are too helpful. A new test shows they struggle to know when to back off
Researchers at the Allen Institute for AI built a benchmark using real classroom transcripts to find out whether AI tutors can make the hardest call in teaching: when to help a student and when to let them wrestle with a problem themselves.
Key points
- Allen Institute for AI released TutorMoments, a benchmark using 462 real one-on-one math tutoring transcripts to test how well AI tutors judge when to help versus when to hold back.
- The dataset includes more than 1,500 decision points flagged by 27 experienced U.S. math teachers across grades 2 to 7.
- Seven large language models, the technology behind chatbots like ChatGPT and Claude, were tested; all of them defaulted to over-helping students when given no specific instructions.
- Giving a model a more detailed prompt improved its scores, but no model consistently matched the judgment human tutors showed at the same moments.
- The benchmark, dataset, and code are all freely available, published in part through the Hugging Face data repository.
A good math tutor, when a student is stuck, usually asks a question rather than answers one. "What do you already know about this problem?" That pause, that small push back onto the student, is deliberate. It is how learning sticks.
AI chatbots are built to do the opposite. Their entire purpose is to be helpful, which in a tutoring context often means explaining the concept, laying out the steps, and walking the student straight to the answer. That short-circuits the effortful thinking that researchers say is central to actually understanding something.
A team at the Allen Institute for AI wanted to know how bad the problem really is, and whether it can be fixed.
How does TutorMoments actually work?
The researchers built a benchmark called TutorMoments around real tutoring sessions, not invented scenarios. They collected 462 transcripts from a U.S. tutoring programme serving mostly low-income schools, stripped out all identifying details, and handed the transcripts to 27 experienced math teachers.
The teachers read through each transcript and marked the moments where a tutor faced a genuine choice: make the problem easier to get into, or push the student to reason harder. Those flagged moments became the test.
At each flagged point, a large language model takes over the session for five exchanges, with a second language model playing the student. A scoring system then checks whether the AI tutor made the right call for that moment, based on what the teachers said the student actually needed.
The three things being measured are whether the model gave appropriate support when needed, pushed for deeper thinking when the student was ready, and avoided giving more help than the moment called for.
What did the AI tutors actually do?
Every model tested defaulted to over-helping. Given a simple instruction to "tutor well," all seven models leaned toward explaining and guiding rather than pushing students to think.
When researchers rewrote the prompt to spell out the trade-off explicitly, scores improved across the board. But no model reliably made the right call the way an experienced teacher does.
The results below show scores between 0 and 1, where 1 means the model made the right call at every relevant moment.
| Prompt type | Appropriate scaffolding | Appropriate rigor | Avoids over-scaffolding |
|---|---|---|---|
| Human tutors (reference) | 0.458 | 0.182 | 0.496 |
| Plain prompt (AI average) | Below human | Near zero | Low |
| Evaluation-aware prompt (AI best) | Above human | Higher | Higher |
One important caveat: the human tutor scores look low because the dataset was deliberately built around missed opportunities, moments where teachers felt something better could have been done. So the human numbers are not a picture of ideal teaching.
What does this mean for students using AI tutoring tools?
If your child or student is using an AI tutoring app, the app is probably doing too much of the thinking for them. That feels helpful in the moment. Over time, it may not be.
The researchers are not saying AI tutors are useless. They are saying the field needs better ways to measure and train this specific skill. A model that never gives away the answer is not automatically a good tutor; what matters is whether it reads the student correctly in each individual moment.
The full dataset, benchmark code, and a technical report are publicly available, as is the dataset on Hugging Face, so other researchers and companies building AI tutoring tools can test their own models against the same real-world decision points.
What to watch for if you use an AI tutoring tool: notice whether it asks you questions or just answers them. A tool that always explains is not tutoring. It is doing the homework.
