The question before the prompt
Three studies of AI that asks instead of answering, two of them from the same journal issue last month, and what they can and cannot say about the question a student would have had to form for themselves.

Correction, 21 November. The version of this edition first published on LinkedIn on 5 November described the Norwegian chatbot study as a short-story assignment and said the pupils who did well had decided what their draft was about before the chat began. The pupils were describing their hometown, the paper does not say that about their topics, and the authors say their patterns shifted within a session. It also described the Welsen coaching tool as limited to asking questions and the Swedish chatbot as nearly identical to it, and called pedagogical friction my own phrase. I have rewritten those sections here and removed the claims I could not trace to the papers. The three levers at the end are my own ideas and are marked that way.
On 20 October, in a short post with no fanfare, Jace Hargis flagged a paper that had come out five days earlier in Education Sciences. The author, Welsen, studied an AI coaching approach, in a course at the University of Nottingham Ningbo China, that asked students about their learning difficulties before it offered any solutions. Of 74 participating students, 31 submitted reflective portfolios, which were the main data, alongside 53 survey responses and 68 weekly learning logs from 12 students. As Hargis summarises it, students said the questions helped them name specific barriers. But 77.4 percent of survey respondents said they sometimes struggled to turn reflection into action, and the same share described some of the AI interactions as generic or repetitive. Welsen cautions that a qualitative design with no comparison group rules out causal claims (Welsen, Education Sciences 16(10): 1728, 15 October 2026; reported by Jace Hargis, 20 October 2026). I have read Hargis's summary, not the paper, which I could not open.
So the AI asked, and the students answered. The step the asking was meant to lead to, turning reflection into action, was where many of them said they struggled.
If every question a student has gets answered, the question the student almost had never gets written down.
A design that refuses to answer
The design that comes closest to refusing sits in the same October issue of Education Sciences. Bradley, Karlén, Segergren, von Feilitzen and Johansson studied a customised ChatGPT chatbot in an eight-week faculty professional-development course at the University of Gothenburg. Apart from administrative questions, it deliberately avoided giving completed answers or summaries and pushed participants back to the course materials. Eleven teachers who used it answered the questionnaire. All eleven found it easy to use. Eight initially identified it as useful, and only one of the eleven said they intended to develop a similar tool for their own teaching. The sample is small, voluntary and from one institution, and the researchers measured perceptions, not learning (Bradley et al., Education Sciences 16(10): 1722, October 2026; reported by Jace Hargis, 21 October 2026).
The researchers call the tension pedagogical friction, as Hargis reports it: the distance between what users expect AI to do and what an educationally designed system asks of them. A system working as designed will withhold the thing a user came for, and the user may experience that as a defect. The usability score and the pedagogical score can pull in opposite directions. A tool that gives you an answer in three minutes is easy. A tool that asks a question that lands on the weak spot in your draft, and keeps asking until you work it, is not.
What the question-only shape makes possible
Earlier in the year, a Norwegian team sat 17 lower-secondary pupils (ages 13 to 14) down for a 45-minute prewriting session with a chatbot built on GPT-4o. The task was to describe their hometown to someone who had never been there. The chatbot asked questions and would not write any of the text. Westbye, Eriksen and Blikstad-Balas, writing in Frontiers in Education in February, analysed 201 question-and-response pairs across 17 dialogues and described three patterns. Some pupils went along with the chatbot's questions and built on them. Others pushed back on or redirected a question, then re-engaged once the focus suited them. In the third pattern the dialogue stalled: two pupils accepted the questions but added almost nothing, and the authors read that as stalled interaction, not resistance (Westbye, Eriksen and Blikstad-Balas, Frontiers in Education 11, 9 February 2026).
The paper does not say how many pupils fell into each pattern, and the authors are plain that the patterns "were not fixed dispositions": the same pupil's dialogue could shift within a session. They are careful about the claim too. They present it as an analysis of dialogue under constraint, not as evidence that the approach teaches better, and it covers one session of one kind of writing. What they do say is that some pupils need additional support and motivation from the teacher or peers in sessions like this. So my takeaway is narrower than the one I first wrote. A tool that only asks makes it visible, inside the session, who is stalling. That is a reason to check in, not a verdict on the pupil.
The usual complaint about students and AI is that they cannot write good prompts. I wonder whether the gap sits earlier. That is my own reading, not a finding of these studies, which were small and were not built to test it. A student who writes a weak prompt has at least written a prompt. The question they would have had to form for themselves, out loud or on paper, before opening a chat window, is the one an always-answering tool never asks for. If the AI's answer stands in for the formulation of the problem, I suspect it removes the part that teaches. I cannot show that here.
Where silence fits in
Writing in AI & Society in August, Humaira Mariyam B. calls the environment this creates in a classroom the "always-answering classroom." A space in which uncertainty keeps getting resolved before it does any work, and the pause after a hard question never gets long enough to become one. The paper is a reflection rather than a trial, and I treat it as such; it argues that silence was doing work, and that erasing it changes what learning is (Humaira Mariyam B., AI & Society, 13 August 2026). The studies above describe what students did with a tool built not to answer. None of them, as far as I could read, tests whether that work lasts longer than work done with a tool that does answer, and the Norwegian authors say outright that theirs is not evidence of instructional effectiveness. I will mark that as a limit rather than paper over it.
Three levers for an assignment this week
Three design moves drawn from the studies above, applied one at a time inside a single assignment. Mine, drawn from reading the papers together, and untested as a set. Edit them to fit the setting.
- Make the student write the question before they open the chat window. One sentence, in their own hand, naming the specific thing they do not yet understand. I cannot point to a study showing this works. It is the move that follows if my reading above is right, and it costs a minute.
- Require at least one session where the AI is a question-only tool. Westbye and colleagues used a chatbot that asked questions and refused to write text; a classroom version can be approximated with a persona prompt that holds for one assignment phase. In their sessions the stalls showed up inside the dialogue, which is early enough to notice. Their patterns also shifted within a session, so treat a stall as a prompt to check in, not a verdict.
- Name a route to a human the student has access to, inside the assignment itself. The Norwegian paper's abstract says some pupils need additional support and motivation from the teacher or peers in sessions like this. A coaching chatbot with no visible exit to a person is a loop a tired student may exit by closing the tab. The route can be as small as a writing-centre coordinator or a scheduled office hour. What it needs to be is named in writing, inside the brief, where the student will see it before they get stuck.
The common move across the three is small. Each one adds a moment where the student puts something on paper before the AI gets to put anything there for them.
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.
One question I would like answered, and I do not have a good one of my own. If you teach, or support people who do, where in your course design is the moment when a student has to write out the question they are stuck on, in their own words, before they are allowed to go anywhere else? And when it is not there, what tends to happen in its place?
Next edition: the second subject in this short arc on where learning hides inside the activities everyone already does. The study groups that still work, and the ones that only look like they are working.
These editions run alongside Learning to Learn With AI, a longer series working the same ground in order.
Key takeaways
- Welsen's questioning-first AI coaching, as Hargis summarises it (74 students, 31 portfolios, 53 survey responses, 68 weekly logs), helped students name specific barriers, but 77.4 percent said they sometimes struggled to turn reflection into action. The design cannot support causal claims.
- In a Gothenburg faculty course, all 11 teachers who used a chatbot designed to avoid completed answers found it easy to use, 8 initially identified it as useful, and 1 intended to build a similar tool. The researchers call the tension pedagogical friction.
- In a Norwegian study of 17 pupils and 201 question-and-response pairs, the authors described three patterns that shifted within a session. They say it is not evidence that the approach teaches better, and that some pupils need extra support from the teacher or peers.
Common questions
What is pedagogical friction? It's the term the Gothenburg researchers use, as Hargis reports it, for the tension between what users expect AI to do and what an educationally designed system asks of them. A chatbot that withholds immediate answers can look less useful even when it is meant to prompt deeper thinking.
What did the Norwegian question-only chatbot study find? Seventeen pupils aged 13 to 14 used a chatbot that asked questions and would not write text, while describing their hometown. The authors described three patterns: pupils who built on the questions, pupils who pushed back and then re-engaged, and a stalled pattern. The patterns shifted within a session, and the authors say the study is not evidence that the approach teaches better.
What can I change in an assignment this week? Three levers of my own, untested as a set: have students write the question they're stuck on before they open the chat, require one session where the AI only asks questions, and name a route to a human inside the assignment brief.
References
- Welsen. (2026, 15 October). From Answer Generator to Questioning Coach: Designing AI-Assisted Coaching for Self-Regulated Learning in Higher Education. Education Sciences, 16(10), 1728. https://doi.org/10.3390/educsci16101728 Summary read via Hargis; the paper itself was not accessed.
- Hargis, J. (2026, 20 October). AI and self-regulated learning. Jace SoTL Substack. https://jacehargis.substack.com/p/ai-and-self-regulated-learning
- Bradley, L., Karlén, Y., Segergren, C., von Feilitzen, C., & Johansson, S. (2026, October). Beyond Answer Generation: University Teachers' Experiences with a Reflective GenAI Chatbot for Professional Learning. Education Sciences, 16(10), 1722. https://doi.org/10.3390/educsci16101722 Summary read via Hargis; the paper itself was not accessed.
- Hargis, J. (2026, 21 October). AI chatbot for PD. Jace SoTL Substack. https://jacehargis.substack.com/p/ai-chatbot-for-pd
- Westbye, A. K., Eriksen, H., & Blikstad-Balas, M. (2026, 9 February). When AI only asks: how question-driven dialogue shapes prewriting in the classroom. Frontiers in Education, 11. https://doi.org/10.3389/feduc.2026.1740044 Task: describe your hometown (adapted from the Norwegian national 8th-grade writing assessment). GPT-4o per Methods. The authors state the study is not evidence of instructional effectiveness.
- Humaira Mariyam B. (2026, 13 August). Reflections on the Always-Answering Classroom: AI, Uncertainty, and Pedagogical Silence. AI & Society. https://doi.org/10.1007/s00146-026-03216-1