Blog · Teaching
How to use AI in the classroom without losing the learning
Use AI so that students still do the thinking. Keep a simple order of work: the student tries first, help from AI, a teacher or a classmate comes next, and later the student shows without help what they can do. Set tools to give hints rather than answers, check learning with short quizzes and questions about students' own work with the machine closed, and do not treat an AI detector score as proof of cheating.
When I taught in an elementary school in Pakistan, the threat to homework was a printed guidebook sold near the school gate, with a ready answer for every exercise. Today the shortcut is a chatbot: free, private, instant and fluent. My book Teaching in the AI Era starts from that change and reaches a plain conclusion: "AI changed what students can hand in far more than it changed what learning requires." This post pulls together the practical answer the book gives to the question teachers ask most: how do you use AI in the classroom without losing the learning?
What has AI actually changed in the classroom?
Chapter 1, "The Day the Homework Changed", names four changes. Help became cheap, detection became unreliable, the teacher's authority shifted, and a finished task stopped being good evidence of learning.
The last change matters most. In a large randomised trial in a Turkish high school, nearly 1,000 students practised mathematics with a standard GPT-4 chatbot, a tutor version with teacher-designed hints, or no AI (Bastani et al., 2025). With the standard chatbot, practice scores rose by 48%. When the chatbot was removed for the exam, those students scored 17% lower than classmates who never had it. The tutor version raised practice scores by 127% and largely removed the harm. Students who leaned on the chatbot did not notice any loss. They felt they had learned.
Students are not simply outsourcing everything, though. In one survey, 83% said AI should never write a whole assignment, and the most common use was to explain a concept (Chen et al., 2026). Yet more than 80% of American students said their teachers had never explicitly taught them how to use AI for schoolwork (Doss et al., 2025). That gap is where a teacher can start.
Why is getting the answer not the same as learning?
Chapter 5, "Learning When Answers Are Free", draws on decades of research that came before chatbots. In a classic experiment, students who reread a passage looked better after five minutes, but a week later the students who had practised recalling it remembered 56% of the ideas against 42% (Roediger & Karpicke, 2006). Effort spent retrieving and producing answers builds memory. An answer handed over before the student has tried skips that effort.
Even tools built to hold answers back depend on students wanting to think. In a trial in Tennessee middle schools, the median student messaged the Khanmigo tutor in only 17% of the practice sessions in which they had made a mistake (Oreopoulos & Low, 2026). One student typed "just give me it". As the book puts it, "A tool can be built to hold back answers, but it cannot make a student want to think."
What order of work protects thinking?
The book's central routine is simple. The student tries first, even badly. Help comes next, as a hint from a tool, a teacher or a classmate. Later, the student shows without help what they can do. It works with any source of answers: a chatbot, a tutor, the back of a textbook or an older sibling. Chapter 5 turns it into classroom habits:
- Attempt first, then ask. Before anyone opens an AI tool, students write down a first step, a guess or a plan, even a wrong one.
- Hints, not answers. Set the class tool to give one hint at a time. Try it as a student first, including a blunt request for the answer.
- A quiz with the machine closed. Once a week, five short questions from memory, with no devices and no marks recorded.
- Ask about their own work. Pick one sentence or step from each submission and ask the student to explain it at the desk.
- Drafts in view. Students show a draft before the final version, and you comment on the thinking as well as the text.
- Agree the line together. The class agrees which uses of AI help learning and which replace it, and the list goes on the wall.
- Name the feeling. Tell students that effortful learning often feels worse even while it works better.
The book is honest that nobody has tested these routines as a package. They draw on the research, and each one is cheap to try.
Should teachers use AI detectors?
The book's advice is no, not as proof. A test of 14 detection tools found that none reached 80% accuracy (Weber-Wulff et al., 2023). In another study, seven detectors wrongly flagged essays by non-native English writers as machine-written in 61% of cases on average (Liang et al., 2023). Teachers do not do much better by eye: in one study, trainee teachers spotted 45.1% of AI texts and experienced teachers 37.8% (Fleckenstein et al., 2024). A lecturer in my doctoral research described a student who made AI text "AI-free" with concealing software in fifteen minutes.
Treat a suspicion as a reason to talk with the student. Ask them to explain a choice and how the work took shape. Keep two decisions apart: what you can say about the student's learning, and whether anyone broke a rule. Uncertain understanding does not prove misconduct.
How can assessment stay fair when AI can write the work?
Chapter 6, "Fair Assessment in the AI Era", starts with validity: does the evidence support the claim you want to make about a learner? A polished essay written at home may show a good text. It cannot, on its own, show that its author can build that argument alone.
So pair open work with a protected check. Students can use approved tools on a project, and then explain a new source, graph or changed constraint in class without help. In a Toronto physical education course, students used AI to help plan a warm-up, then had to "lead and explain the routine in person". A zoologist in my study moved marks from assignments to quizzes and added a short viva on each assignment. Report the two claims separately, so a single total does not hide the gap between doing well with support and understanding alone.
What if my school has one shared phone and a weak signal?
Chapter 7, "Access, Equity and AI", is written for these classrooms. With one teacher phone, prepare while the connection works: save a checked passage, diagram and questions before school, then copy them onto the board or use a printed master. Students work with the material together, and their learning does not depend on taking turns with your phone. If the connection fails mid-lesson, use a stopping rule: one brief attempt to reconnect, then move to the prepared paper alternative. The book is clear that no trial has yet tested one teacher phone supplying AI to a whole low-resource class, so treat this as careful practice, not proven method.
The good news is that the routines that protect thinking need very little equipment. In the book's words, they "need little more than paper and a teacher's attention." Attempting first, quizzing from memory, explaining one's own work and showing drafts all work with chalk.
Where can I read the book?
Teaching in the AI Era is free to read online, chapter by chapter, with reflection questions for a department meeting at the end of each one. If you would like your own copy, the ebook is at books2read.com/u/mdrgNO, and the paperback is on Amazon.
Questions people ask
How should teachers use AI in the classroom?
Keep the student doing the thinking. Ask students to attempt a task first, let AI or a teacher give a hint next, and later check what each student can do without help, for example with a short quiz with devices closed.
Does AI harm student learning?
It can. In a randomised trial in a Turkish high school, students who practised with a plain GPT-4 chatbot scored 17% lower on the exam without it than classmates who never had it. A tutor version that gave hints and withheld answers largely removed that harm (Bastani et al., 2025).
Are AI detectors reliable enough to prove cheating?
No. A test of 14 tools found none reaching 80% accuracy, and detectors have wrongly flagged non-native English writers. Treat a concern as a reason to talk with the student, not as proof.
Can I use AI in a classroom with one phone and poor internet?
Yes, with care. Prepare and check materials while the connection works, then use the board or paper in the lesson. The routines that protect thinking, such as attempting first and quizzing from memory, need no device at all.
About the author. Dr Muhammad Tayyab Ilyas is the author of Teaching in the AI Era, a book for teachers on AI in schools and universities. He taught science and mathematics in a government elementary school in Gujranwala, worked as the district's assistant education officer and teacher trainer, and holds a PhD in Education from the Universitat Autònoma de Barcelona.