Teaching a kid to ask a good question sounds like the softest skill in the room. It’s also been studied carefully for decades, at least for a related skill: teaching a reader to generate their own questions about a passage, using nothing more than who, what, where, when, why and how, raises reading comprehension. Barak Rosenshine and colleagues pooled these training studies in 1996: a median effect size of 0.36 on standardized tests, and a much bigger 0.86 on tests the researchers wrote themselves, which is the friendlier yardstick and worth naming as such. Among the most effective approaches in the studies reviewed were the plainest: hand students signal words or generic question stems, then have them practise filling them in (Rosenshine, Meister, & Chapman, 1996).
That finding is about thirty years old and has nothing to do with artificial intelligence. A two-year trial in 18 Tennessee middle schools gives it a new and immediate use: the median student using an AI maths tutor messaged it in only 17 percent of the practice sessions where they made a mistake, and mostly sent bare answers when they did (Oreopoulos & Low, 2026, https://edworkingpapers.com/ai26-1551, a working paper, not yet peer reviewed, covered here last edition).
Key takeaways
- Students mostly send AI tutors bare answers or “how do I” questions. Research described the same help-seeking problem two decades before generative AI.
- Teaching students to generate their own questions has decades of evidence behind it in reading comprehension.
- Studies of teaching students to prompt an AI are small and early. An explicit finish line in the tutor’s setup came with a smaller shortfall.
The old problem, in a new box
Software that offers help on request isn’t new, and neither is students declining to ask for it. Vincent Aleven and colleagues reviewed the research on what they called interactive learning environments, the pre-AI ancestors of an AI tutor, and reached a conclusion this edition is really just applying to a newer tool: effective help-seeking predicts better learning, and most students don’t do it effectively, for reasons that sit as much in how the help is designed as in the student (Aleven, Stahl, Schworm, Fischer, & Wallace, 2003).
In a classroom study of a tutoring agent called the Help Tutor, Aleven and a different team tried fixing it directly: real-time feedback coached students toward asking at the right moment instead of clicking through hints or avoiding help altogether. In plain terms, here’s what they found: coaching students to ask for help more deliberately changed their behaviour, and the change held up after the coaching stopped. It didn’t change what they learned (Aleven, Roll, McLaren, & Koedinger, 2016).
“Coaching students to ask for help more deliberately changed their behaviour. It didn’t change what they learned.”
Honestly, that’s the finding I’d have cut if I were selling you something. Sit with it before getting hopeful about the rest of this edition. It also points at a gap worth naming plainly: the studies Rosenshine reviewed trained a different skill (writing questions about a text) than Aleven measured (asking a tutor for help). Nobody has tested whether the old training transfers to the new context. That’s why the rest of this edition is a set of small, separate tests.
What students are actually typing
A 2026 preprint, not yet peer reviewed, analysed 6,113 real messages students sent to Large language model (LLM) chatbots while studying, across ordinary practice and graded coursework. It found the same shape as the Tennessee trial: procedural questions, “how do I,” dominate, and dominate even more once a grade is on the line (Neagu, Messer, Johnson, & Nelson, 2026). Bare answers and “how do I” are close cousins. Neither is the kind of question that pushes back on a student’s own thinking, which is closer to what Rosenshine’s signal words were built to force, on a different task, decades before any of this software existed.
Teaching the AI-specific version of an old skill
One small 2025 preprint tests something closer to the actual gap: a short lesson in what its authors call pedagogical prompting, phrasing a request so the AI answers with guidance instead of a solution. Ruiwei Xiao and colleagues taught it to undergraduate computer-science students in a scenario-based lesson, then measured them before and after. Twenty-two students, about the size of one classroom, is a pilot with no comparison group, and college students at that (Xiao et al., 2025). But for a reader who teaches or designs at university level, that’s the most relevant sample in this edition: a short, deliberate lesson was followed by an improvement in the AI-directed help-seeking skill it targeted, and by students saying they’d keep using it. That’s an improvement in the asking, the same category Aleven’s 2016 result just warned isn’t the same as an improvement in the learning.
The lever a teacher already has
There’s a second way in, one that doesn’t depend on training the student at all. A 2026 preprint, also not yet peer reviewed, studied 39 classrooms across 20 teachers, looking at what happened when teachers pre-authored the AI’s side of the conversation, the setup instructions and starter prompts a student meets on arrival, instead of leaving the dialogue to whatever the student typed first. Of 1,479 resulting conversations, 71 percent stayed fully aligned with what the teacher intended, and fewer than 1 percent went seriously off track. The gap: 38 percent of conversations still landed at a shallower level of thinking than the teacher had targeted, worse at the hardest targets, closer to half when teachers were aiming for real analysis over recall. One thing tracked with a smaller gap: prompts with an explicit finish line, a plain statement of what “done” looks like, came with a shortfall about a fifth of a level smaller (Liu, Sun, Esbenshade, Tian, Zhang, & He, 2026). That’s an association from an observational study, not a controlled test.
Put the two together and there are two separate levers here: teach the student the question, and set the prompt so a weaker question still lands somewhere useful anyway. Both fit inside a teacher’s actual week, or a planning period spent building a tutor’s setup, which is exactly the kind of work a faculty developer or instructional designer does for a living.
| Study | Type and sample | What it tested | What it found |
|---|---|---|---|
| Rosenshine, Meister and Chapman (1996) | Review of training studies | Teaching students to generate questions about a passage | Median effect 0.36 on standardized tests, 0.86 on researcher-written tests |
| Aleven et al. (2003) | Review of interactive learning environments | Help-seeking in pre-AI tutoring software | Effective help-seeking predicts learning; most students don’t do it effectively |
| Aleven, Roll, McLaren and Koedinger (2016) | Classroom study of the Help Tutor | Coaching students to ask for help at the right moment | Behaviour changed and held; what they learned did not change |
| Neagu et al. (2026) | Preprint, not peer reviewed; 6,113 student messages | What students type to chatbots while studying | Procedural “how do I” questions dominate, more so when graded |
| Xiao et al. (2025) | Small preprint; 22 undergraduates, no comparison group | A short lesson in pedagogical prompting | Improved AI-directed help-seeking; students said they would keep using it |
| Liu et al. (2026) | Preprint, observational; 39 classrooms, 20 teachers, 1,479 conversations | Teacher-authored setup prompts | 71% fully aligned; 38% still shallower than intended; explicit finish line tracked with a smaller gap |
Four things to try this week
- Give them the signal words, as a hypothesis worth testing yourself. The studies Rosenshine reviewed often used six concrete prompts: who, what, where, when, why, how. Nobody has tested them on a chatbot yet, so treat this as your own small experiment: put the six words next to the AI tutor and ask for one of them in a student’s first message when they’re stuck.
- Show a weak prompt and a better one, side by side (a version of edition 8’s “teach the question” tip, worth repeating now that Neagu’s data explains why). Neagu’s data says “how do I” is the default question. Don’t just assume students will improve on their own; put an actual weak prompt on the board, then a better version, so the improvement is visible.
- State the finish line before the student ever opens the tool. This is the one item here built for whoever configures the tutor, not just the classroom teacher. Liu’s team found a smaller gap when the setup stated plainly what a finished answer looks like. That’s lesson planning, and it’s the one item on this list I’d try first: it costs a sentence.
- Watch actual behaviour before you trust it. Aleven’s 2016 result is the guardrail on this whole edition: better asking habits can show up before better scores do, or without them ever arriving. Check work with the tool closed, the same habit worth carrying into any AI-tutoring trial.
If you’ve taught a class to ask better questions of a book or a teacher: did a version of Rosenshine’s thirty-year-old trick actually carry over when the thing being asked was a chatbot, or did it need reteaching from scratch for that context? I don’t have a clean answer, and nothing in the research above tests it directly.
How this was made. Full disclosure, and I mean it as a courtesy, not a confession. AI tools help me research, check my claims, tidy my grammar, and stay on top of far more reading than one person can. The thinking and the judgment are mine, I edit every word into my own voice by hand, and then I put my name on it. Telling you that is the whole point of keeping it human.
These editions run alongside Learning to Learn With AI, a longer series working the same ground in order.
Common questions
How do you teach students to ask better questions? Model it. Show a weak question beside a better one, practise together, and give students a starting set of question words. Training in question generation improved reading comprehension in the studies Rosenshine and colleagues pooled, with a median effect size of 0.36 on standardized tests.
Why do students send AI tutors bare answers? Asking well is a skill most students were never taught, and the problem predates AI. Research on interactive learning found that effective help-seeking predicts learning, and that most students don’t do it well. Under grading pressure, “how do I” questions dominate even more.
References
- Rosenshine, B., Meister, C., & Chapman, S. (1996). Teaching students to generate questions: A review of the intervention studies. Review of Educational Research, 66(2), 181-221. https://journals.sagepub.com/doi/10.3102/00346543066002181
- Aleven, V., Stahl, E., Schworm, S., Fischer, F., & Wallace, R. (2003). Help seeking and help design in interactive learning environments. Review of Educational Research, 73(3), 277-320. https://journals.sagepub.com/doi/10.3102/00346543073003277
- Aleven, V., Roll, I., McLaren, B. M., & Koedinger, K. R. (2016). Help helps, but only so much: Research on help seeking with intelligent tutoring systems. International Journal of Artificial Intelligence in Education, 26(1), 205-223. https://link.springer.com/article/10.1007/s40593-015-0089-1
- Oreopoulos, P., & Low, N. (2026). One click away: AI tutoring with Khanmigo in a two-year school experiment [Working paper, not yet peer reviewed] (EdWorkingPaper 26-1551). Annenberg Institute at Brown University. https://edworkingpapers.com/ai26-1551 (Return from edition 8.)
- Neagu, A., Messer, M., Johnson, P., & Nelson, R. (2026). “How do I …?”: Procedural questions predominate student-LLM chatbot conversations [Preprint]. arXiv:2602.18372. https://arxiv.org/abs/2602.18372
- Xiao, R., Hou, X., Ye, R., Kazemitabaar, M., Diana, N., Liut, M., & Stamper, J. (2025). Improving student-AI interaction through pedagogical prompting: An example in computer science education [Preprint]. arXiv:2506.19107. https://arxiv.org/abs/2506.19107
- Liu, A., Sun, M., Esbenshade, L., Tian, V., Zhang, Z., & He, K. (2026). Teacher-authored prompts for configuring student-AI dialogue: K-12 classroom implementation [Preprint]. arXiv:2604.16738. https://arxiv.org/abs/2604.16738