We can now upload a photograph of a difficult problem and have AI produce a solution. Being able to ask immediately about a sticking point is appealing, but an answer appearing on screen and becoming able to solve the problem yourself are different things. If homework gets finished faster but a similar problem stops you again the next day, your study method needs attention. When choosing educational AI, look at what the learner must do, as well as how fluent the answers sound.
This article proposes a study process based on public research checked on October 25, 2026. The fraction problems and learning records below are fictional examples explaining the principles, not product performance tests or records of real students. The representative image is also an illustrative photograph of students studying with a laptop and books in a library. It does not document participants in the studies or prove that AI is in use.
1. Separate assisted results from independent ability
To evaluate a learning tool, distinguish two outcomes first. One is how well you complete tasks while AI is available; the other is whether you can explain the concept and solve a new problem after putting the tool away. A finished answer can demonstrate the first without automatically establishing the second. Nodding along while reading an explanation does not replace a check of understanding.
Bastani and colleagues' 2025 PNAS study compared a general chat-style AI with an AI tutor incorporating teacher-prepared hints and solution information in high-school mathematics in Türkiye. Performance during assisted practice differed from subsequent unaided assessment. The general-chat group performed worse on that assessment than a comparison group without AI. With the hint-oriented tutor, the negative difference was largely reduced, but there was not significant evidence of better independent assessment performance than the comparison group.
This does not mean that AI necessarily reduces ability. The findings come from a particular school, subject and set of learning activities. They do, however, suggest a useful question for study design: did you solve the practice problem by receiving an answer, or build an understanding you can use yourself? Check these separately.
2. Read positive studies alongside the tutor design
Kestin and colleagues' 2025 Scientific Reports study compared a structured AI tutor with classroom active learning across two introductory university physics lessons. The 194 students included in the analysis experienced both methods in turn, and the AI-tutor condition produced greater gains on the assessment immediately after the lesson. This was an educational environment with teacher-prepared, question-specific explanations and solutions, staged progression and feedback, rather than simply a general chat window.
Treating the two studies as votes for and against AI obscures their differences. Learner age, subject, prior knowledge, practice, comparison conditions and assessment timing differed. Improvement after a short lesson cannot be extended to long-term retention or all ages. Before relying on a product's AI label, ask when answers appear, who checks inaccurate feedback and whether there is a period for working alone.
3. Write one learning goal before entering a question
The same tool calls for different use depending on the goal. Instead of finishing ten fraction problems, write the ability you want to develop today: explaining why unlike fractions must be expressed in equal-sized units before addition. In programming, the goal might be explaining a loop's stopping condition or finding an erroneous line, rather than receiving completed code.
Next, put what you currently know on paper. Separate the problem's conditions, the relevant principle, your attempted solution and the sticking point. This makes the scope of AI assistance smaller. Asking for a complete solution from the start mixes what you understand with what AI has done for you. A short first attempt gives information about what you do not know, not just whether your answer is right.
You can specify the amount of help in your question: identify only the first incorrect step in my solution and do not reveal the answer yet. Such requests are not always followed. If AI displays the full solution immediately, manage your own next action: cover the window and write what needs changing in the original attempt instead of copying the answer.
4. Take hints one step at a time and explain in your own words
Start with a small hint. Ask first what information to look for in the problem, then request a relevant concept if you are still stuck. Next, obtain feedback on a particular line of your work, and finally look at as much of an example as needed. This is a practical starting point for avoiding a full answer too early, not a claim that this sequence is optimal for everyone.
What matters is the action after the hint. Pause the chat and revise your own solution. Instead of saying that you understand, write one sentence explaining why the change is necessary. Difficulty explaining may signal an unfamiliar term or a skipped step. Narrow the next question to that point before asking for a longer answer.
If a completely unfamiliar concept prevents you from starting, reducing hints at all costs is not helpful. Study the basic textbook explanation or a teacher's example first, then attempt a similar new problem yourself. Minimizing assistance is not the goal. Distinguishing explanations needed to begin understanding from processes you need to practise yourself is the goal.
5. A fictional example: keep part of the fraction work yourself
Suppose a learner writes 4/10 for 3/4 + 1/6. Instead of asking for the answer, they could say: I added the numerators and denominators separately; guide me with one question about whether I am adding unlike units. Feedback must be checkable against a textbook or with a teacher. AI can explain an incorrect rule even in a friendly tone.
The learner draws or writes the conversion into equal-sized pieces. Since 3/4 is 9/12 and 1/6 is 2/12, the sum is 11/12. The key explanation is not to multiply denominators mechanically, but to convert to equal-sized units and then add the number of pieces. Record why the original 4/10 was inappropriate as well as the final calculation.
Then close the screen and solve 2/3 + 1/4. This different example checks whether you remembered only the previous numbers or can apply the reason for a common denominator. Rewriting a solution while its answer remains visible and independently solving a problem with changed conditions must be recorded separately. These numbers and this sequence are illustrative, not a validated instructional programme.
6. Plan checks without the tool and review on another day
At the end of a study session, put AI and your notes aside briefly and complete one explanation or problem. Depending on your goal, explain a core concept in your own words, say why another method is wrong, or solve a problem with changed numbers or circumstances. Record an attempt to recall before looking separately from the feeling of remembering after seeing the answer.
On the next day or at a later review, attempt the work independently before checking only what is needed. This review timing is a practical example, not an optimal interval for every learner and subject. For an incorrect problem, note the reason in a line, such as failing to align units or missing a condition, rather than copying the whole answer. Checking a new problem targeted to that reason makes your next question more specific.
A weak independent result does not mean that all AI use has failed. Examine whether the explanation was too advanced, a prerequisite concept was missing or a hint exposed the answer too soon. If you repeatedly fail alone, changing the study route through a teacher or reliable textbook may help more than extending the same chat indefinitely.
7. Check accuracy and privacy separately
Book titles, paper links, equations and code supplied by AI all need checking. A link opening successfully and its document supporting a claim are different matters. Give priority to textbook definitions and solution criteria; if AI conflicts, locate the original material and ask a teacher. Check the points that can change an answer: units and signs in mathematics, assumptions in science, and input conditions in code.
UNESCO's guidance on generative AI in education emphasizes human-centred educational design, privacy and age-appropriate use. Do not upload documents containing a student's name, school, identification number, grades or contact details unchanged. Remove personal information and use only necessary excerpts. Minors should check guardian and school requirements and the service's current age conditions. Apply the same care to camera and voice input.
School assignments may have separate rules. Check permitted assistance and disclosure requirements, and distinguish your own thinking from AI-assisted parts. Submitting a generated answer as your own also makes learning difficult to assess. Do not share accounts for convenience or upload someone else's work or non-public teaching materials without permission.
8. Record the ability that remains, not just time spent
A learning record can be short: date, today's goal, first attempt, assistance received, result without the tool and the next check. For example: explain common denominators; added numerators and denominators; one hint about units; solved another problem alone; explain the reason again tomorrow. This shows what became possible after which assistance more clearly than chat counts or quantities studied alone.
Read the records periodically and adjust the assistance. If you still need a completed solution for the same concept, divide the goal further or return to basics before looking for a new tool. If you can explain independently and apply the idea to other problems, reduce AI interventions. The comparison should be your observed performance, not someone else's usage time or an advertisement.
Educational AI's value extends beyond displaying an answer quickly. It can help organize questions and locate sticking points while leaving thinking, explaining and checking to the learner. In your next session, write a goal and first attempt before turning it on, and make one unaided check before finishing. That small record starts to distinguish convenient assistance from ability you have actually developed.