Recent education headlines are making the same point from different directions: AI tutors can help learners, but only when they keep students thinking instead of quietly taking over the work. The newest signal comes from reporting on the 2025 PISA cycle, where analysts are examining how teenagers use AI tools for studying, writing, summarizing, and early research. A recent summary of the OECD findings reported a pattern that should make schools cautious: heavier or more mechanical AI use was associated with weaker academic performance, while more purposeful use looked less harmful and sometimes more promising.
That does not mean AI tutors are useless. It means the learning question has shifted. The practical issue is no longer simply, “Can a chatbot answer?” It is, “What kind of interaction makes the learner ask, test, explain, and revise?” The OECD Programme for International Student Assessment is built around how well 15-year-olds can apply knowledge in reading, mathematics, and science. If students use AI to avoid reading, drafting, estimating, or checking their own reasoning, the tool may improve the homework product while weakening the skill PISA and real learning both care about.
What Recent Studies Add
Two 2026 research papers sharpen the point. In “Warning About AI Fallibility Increases Help-Seeking in an Intelligent Tutoring System”, researchers ran a classroom experiment with 252 school students using a math intelligent tutor. One version warned students that the system could make mistakes; the other did not. The warning did not change the underlying tutor, but it changed learner behavior: students who saw the warning asked for more hints.
That is a small design choice with a large learning implication. When students believe an AI system is authoritative, they may stop monitoring it. When they are reminded that the system can be wrong, they become more active users. They ask for support, compare information, and treat the answer as something to examine rather than something to accept. In learning-science language, the warning appears to support better “trust calibration”: learners neither reject the tool nor surrender judgment to it.
A second paper, “Regulating the AI Tutor: Intentions, Help-Seeking, and Self-Regulated Learning in Adolescent GenAI Use”, looked at 98 ninth-grade students using a web-based large-language-model tutor for mathematics preparation. Students often said they wanted scaffolded support before the chat began, but the interaction logs told a more complicated story. Their conversations were dominated by instrumental requests, with little explicit monitoring or evaluation of the AI’s help.
That gap matters. Many learners want help that builds understanding, but under pressure they naturally ask for the shortest path to an answer. A chatbot can make that shortcut feel normal because it is fast, fluent, and always available. The risk is not just cheating. The quieter risk is under-practicing the mental moves that produce durable learning: defining the problem, naming what is confusing, predicting the next step, checking whether an explanation fits, and trying again after feedback.
Why It Matters For Questions
The strongest conclusion from these studies is that question quality is now a core study skill. A weak AI prompt asks the tutor to finish the task. A stronger learning prompt asks the tutor to protect the learner’s effort while making the next step clearer.
For example, “Solve this for me” can produce a neat answer and very little learning. “Ask me one hint at a time, and wait for my attempt before giving the next hint” keeps the learner in the problem. “Here is my explanation; find the first place where my reasoning breaks” forces comparison and self-correction. “Give me a similar problem, but do not show the solution until I try” turns AI from an answer machine into a practice designer.
This is also where assessment news and AI tutor research overlap. If schools only grade finished products, students have an incentive to use AI for polished output. If teachers and learners value drafts, attempts, explanations, error review, and oral defense of reasoning, AI can be used in ways that reveal learning rather than conceal it. The tool then becomes part of a feedback loop instead of a replacement for the loop.
The Practical Learning Conclusion
AI tutoring should be judged by what the learner does next. Good use leaves a student with more accurate self-knowledge, a clearer next attempt, and some memory of the reasoning. Poor use leaves only a completed answer.
- Ask for hints before answers. Start with, “Give me the smallest useful hint,” or, “Ask me a question that helps me find the next step.”
- Show your thinking first. Paste your attempt and ask the AI to identify the earliest error, not to rewrite everything.
- Check understanding immediately. After an explanation, close the chat and explain the idea from memory in two or three sentences.
- Request a transfer problem. Ask for a new example with different numbers, context, or wording so you can test whether you learned the method.
- Keep uncertainty visible. Use prompts such as, “What might be wrong with this answer?” or, “What assumption should I verify?”
The recent evidence does not support panic or blind enthusiasm. It supports a more disciplined habit: use AI tutors to make better questions, not to make effort disappear. The learner’s job is still to struggle productively, check understanding, and make the next attempt. The tutor’s best role is to keep that process moving without stealing it.