Recent AI tutor research is pointing toward a practical rule for learners: ask for help that keeps you thinking, not help that finishes the thinking for you.
The optimistic side of the news is real. A recent Brookings review of generative AI tutoring research notes that rigorous trials are beginning to show useful learning gains when AI systems are designed around tutoring functions such as feedback, dialogue, and guided practice. The review also warns that the evidence base is still developing and that good tutoring depends on design, subject matter, and how the learner interacts with the tool. Read the review at Brookings.
A 2025 study in Scientific Reports adds another encouraging data point. In that experiment, students using a custom AI tutor learned more in less time than students in an in-class active learning comparison, and they reported higher engagement and motivation. The authors emphasized that the tutor was built around research-based teaching practices, not simply around giving quick answers. See the study at Nature Scientific Reports.
But newer findings complicate the easy story that more AI help always means more learning. A 2026 preprint titled “AI Assistance Reduces Persistence and Hurts Independent Performance” reports that participants who used AI assistance during short reasoning tasks performed worse after the assistance was removed, especially when they had asked the AI for direct answers. The paper’s strongest practical warning is about habit formation: even brief answer-focused help can reduce persistence when the learner later has to work alone. The preprint is available at arXiv.
A separate MIT Media Lab study on AI and misinformation points in the same direction. Participants became more accurate while an AI chatbot helped them evaluate news items, but their later unassisted performance declined after weeks of relying on the tool. MIT News summarized the study as evidence that AI can improve immediate accuracy while failing to build the user’s independent judgment. Read MIT’s report at MIT News.
Why It Matters
For learners, the issue is not whether AI tutors are good or bad. The sharper question is what kind of help request turns an AI tutor into practice, and what kind turns it into an answer machine.
Human tutors already know this distinction. A good tutor rarely begins by solving the whole problem. They ask what the learner has tried, point to the next useful step, explain the reason behind a method, or give a similar example before asking the learner to continue. The learner still has to retrieve knowledge, compare options, make errors visible, and repair the solution.
AI tutors can support the same process when learners prompt them well. They can give quick feedback, generate extra practice, explain a confusing step, or offer a worked example for comparison. The risk appears when the learner repeatedly asks for final answers, full drafts, completed solutions, or polished explanations before making a serious attempt. That feels efficient in the moment, but it removes the desirable struggle that builds recall, transfer, and confidence under test conditions.
How Learners Should Ask for Help
- Ask for a hint first when you understand the goal but are stuck on the next step. A good prompt is: “Give me one small hint, but do not solve it.”
- Ask for an explanation after you have tried and can name the confusing point. A useful prompt is: “Explain why this step works and where my reasoning went wrong.”
- Ask for an example when the concept is unfamiliar. Then close the example and solve a new problem without looking.
- Ask for feedback after producing your own answer. Feedback is more useful when there is something real to inspect.
- Switch to independent practice after one or two guided problems. If you keep needing the same hint, make a smaller practice set rather than asking for another full solution.
The Practical Conclusion
The best help-seeking habit is a ladder: attempt, hint, explanation, example, feedback, independent retry. Do not start at the top by asking for the finished answer. Start with your own attempt, ask for the least help that gets you moving, and always end with a no-help problem.
That final independent retry is the part many learners skip, but it is the part that tells you whether the help became learning. If the AI disappears and you can still explain the method, solve a fresh problem, or spot your own error, the tool supported learning. If the AI disappears and the work collapses, the help was too strong too soon.