top of page

How Students Teaching Each Other with AI Is Solving the Tutor Shortage

3 days ago
7 min read

It is the week before finals at a polytechnic in Kuala Lumpur. The tutoring centre has a waitlist. The lecturers are fielding emails at 11 p.m. The students who most need extra time are the ones least likely to ask for it in person.


Down the hall, a group of four second-year students in a business analytics programme have built something. One of them, a quiet woman who barely spoke in tutorials, has spent three evenings structuring her notes on regression analysis into an AI tutor she built herself. Her classmates are using it. The questions it asks are good: specific, scaffolded, calibrated to the assessable content. She built it because preparing it helped her understand the material better. Her peers are using it because it is available at 11 p.m., it does not judge, and it knows exactly what the lecturer is going to test.


No faculty member coordinated this. No budget was allocated. The tutor shortage, for this group, is not acute that week.

The Arithmetic of the Shortage


UNESCO's projection is stark: the world will need 44 million more teachers by 2030 to meet global education demand. In Southeast Asia, student-to-educator ratios are already under pressure. Philippine higher education runs at roughly 23 students per faculty member on average. Indonesian public universities hover near 30 in some faculties. Even Singapore, with its comparatively well-resourced system, faces the compound pressure of growing enrolment, rising student expectation for individualised attention, and educators whose non-teaching workloads have not shrunk.


The traditional response to a tutor shortage has been to hire more tutors. This is expensive, slow, and rarely keeps pace with demand. The more recent response has been to build AI tutors, chatbots deployed by institutions to answer student questions at scale. These help. They are better than nothing. But they carry a design problem: they are built by institutions, for students, in a relationship where the student is still the passive recipient.


Learning science has known for decades that this is not the most efficient configuration.

The Protégé Effect


In 1984, psychologists studying learning retention noticed something counterintuitive: students who were asked to teach a topic to someone else learned it better than students who simply studied it themselves, even when they had the same amount of preparation time. The phenomenon acquired a name, the protégé effect, and the research behind it has only grown more robust in the forty years since.


When you prepare to teach, you identify your own gaps. You have to structure information logically for another person, which forces you to understand it structurally yourself. You generate explanations in your own language, which encodes the knowledge more deeply. And when your student asks a question you cannot answer, you experience an immediate, concrete signal that you need to go back to the material.


Peer tutoring, the structured application of this principle, has been studied extensively. A 2021 meta-analysis published in Educational Psychology Review examined 74 studies across higher education and found that peer-taught students outperformed control groups on assessments in 68% of cases. The effect was stronger in STEM subjects, where procedural knowledge is easier to scaffold. The tutors themselves showed even larger learning gains than the students they taught.


The problem has never been whether peer tutoring works. The problem has been whether it scales.

Why Peer Tutoring Has Always Been Hard to Scale


Traditional peer tutoring programmes require significant institutional infrastructure: recruitment and training of peer tutors, a matching system pairing tutors with appropriate tutees, quality oversight to ensure the tutor is not confidently transmitting misconceptions, and scheduling coordination that assumes both parties have time that aligns.


In practice, this limits peer tutoring to a small fraction of the student population, typically high-achieving students who are recruited and paid as formal tutors, available for booked sessions at the learning centre. Students who need help at 11 p.m. on a Tuesday, who are too anxious to book a session, or whose question is too specific to fit a generic tutoring slot are not well served.


AI changes three of these constraints simultaneously.


The first is availability. An AI tutor does not have a schedule. It does not get tired, does not run out of patience, and does not prioritise the students who are already performing well.


The second is preparation overhead. One of the biggest barriers to peer tutoring at scale is that being a good tutor requires preparation that many students do not have time to do formally. But when the act of structuring your knowledge into an AI is itself the preparation, when the tool makes the process low-friction enough that a student does it naturally, the barrier dissolves.


The third is quality signalling. A well-structured student-built AI tutor, grounded in the course's actual content, produces questions and explanations that reflect that student's genuine mastery. If the student has misunderstood something, that misunderstanding will surface in what the AI produces, and it can be caught and corrected before it propagates.

Students as Builders, Not Just Users


The more transformative shift is not in how students use AI, but in what happens when students build it.


Noodle Factory's NoodleMe is built on exactly this principle: students become teachers when they structure their own AI tutor and share it with classmates. A student uploads their lecture notes, annotated readings, and past-paper responses. The system helps them structure that material into a tutor that can quiz their peers, explain concepts in their voice, and answer questions that stay within the scope of what they have actually understood and documented.


The act of building is the learning. A student who has had to decide which concepts deserve a practice question, which misconceptions are worth flagging, and how to explain a difficult idea clearly to a peer who is not yet there, has done something more cognitively demanding than writing an essay. They have taught.


And the output is useful to others. In a well-connected cohort, the best student-built tutors become shared resources. The student who built the regression analysis tutor in Kuala Lumpur did not intend to become a resource for her cohort. But the quality of what she built, and the accessibility of the tool, meant that her work extended its reach without her having to be present.

What Faculty Actually Gain


The concern that peer-mediated AI learning takes something away from educators is understandable. It is also, the evidence suggests, misplaced.


Educators who have introduced student-built AI tutors into their courses report two consistent effects. The first is a change in the quality of questions they receive in class. When students have already worked through the accessible, well-structured parts of a topic with an AI, the questions that surface in the tutorial room are more interesting: more conceptual, more edge-case, more genuinely challenging. The lecturer's time shifts toward the questions that most benefit from a human expert.


The second is an insight into student understanding that traditional assessment does not reveal. A student-built AI is a kind of externalised knowledge map. Where the knowledge is clear, the AI's outputs are coherent. Where it is absent or distorted, the AI reveals it. Lecturers who have access to student-built tutors can identify conceptual gaps at the cohort level, not just the individual level, before the examination reveals them.


Dr. Alison King, whose foundational research on peer learning at California State University shaped much of what we now know about structured peer tutoring, argued in 1993 that the classroom should move "from sage on the stage to guide on the side." Thirty years later, AI has finally given educators the infrastructure to make that shift practical, not as a metaphor but as a design principle for how learning happens.

The Coordination Problem, Revisited


The remaining challenge is coordination. Student-built AI tutors are only as valuable as the network through which they circulate. A tutor built by a high-achieving student that stays on her laptop helps no one but her. The institutions that are seeing the most benefit from student-led AI learning are those that have built light-touch systems for sharing: shared libraries, course-linked repositories, or peer review mechanisms that let the best student-built resources surface and be used by the cohort.


This does not require heavy institutional infrastructure. It requires a norm: that building and sharing a good AI tutor is valued, recognised, and in some cases formally credited. Some institutions are beginning to treat peer knowledge-sharing, including AI-mediated sharing, as a form of academic contribution. At Singapore Management University, peer teaching activities have been incorporated into some programme rubrics as assessable evidence of mastery.


The norm shift matters more than the technology. Students already know, intuitively, that explaining something to a friend is one of the best ways to understand it. What AI adds is the ability to do that at scale, asynchronously, and in a form that persists beyond the conversation.

The Night Before the Exam


It is 11 p.m. The tutoring centre has closed. The lecturer is not checking email. The student who needs to understand conditional probability before tomorrow morning has three options: give up, stare at the textbook until her eyes stop working, or open the AI tutor her classmate built last week and start working through the questions.


The classmate who built that tutor is sleeping better than she has in a month. She built it three weeks ago, in the week before her own understanding finally clicked. She did not build it to help others. She built it because structuring her knowledge for a hypothetical student forced her to figure out what she actually understood.


Both of them, it turns out, needed the same thing: a system that treats learning as something you do for others, not just for yourself.


That system has existed in learning science for decades. AI has finally made it cheap enough, accessible enough, and low-friction enough to scale. The tutor shortage is not solved. But for the students working together at 11 p.m., it is, for tonight, not the binding constraint.

Further Reading


Sources


bottom of page
Chat with Walter