AI Can Answer a Question, But Can It Teach?

    22 September 2026
    10 min read
    Listen to this post
    AI Can Answer a Question, But Can It Teach?

    By Julian, CEO & Co-founder of SideCog


    The debate about AI in education is often reduced to a choice between excitement and fear. The more useful question is whether a tool is designed to complete a task or to help a child learn.


    A child can now type almost any homework question into a general-purpose AI tool and receive a polished answer in seconds. That is impressive. From a teaching point of view, it is also the least interesting thing AI can do.


    An answer is not the same as learning. A completed task is not proof that anything has changed in long-term memory. A child may appear to make rapid progress with AI beside them and still be unable to explain, retrieve or apply the idea when the screen closes.


    The headlines are asking the wrong question


    One of the most widely shared education headlines this month said that students who use AI generally score worse at school. It drew on the OECD PISA 2025 study of more than 760,000 15-year-olds across 91 countries and economies. Students who reported using AI for school tasks such as summarising texts, drafting work or carrying out research scored around 20 points lower in science than non-users on average, roughly the equivalent of a year of schooling.


    That finding deserves attention, but it does not prove that AI caused the lower scores. The data also showed a more hopeful pattern. Moderate, intentional use for learning was associated with better outcomes than daily dependence, especially where students had been taught to judge AI output critically. The important distinction is not simply between using AI and avoiding it. It is between AI that replaces thought and AI that supports it.


    The OECD Digital Education Outlook 2026 makes the same point more directly. General-purpose AI can improve the quality of a piece of work without producing a genuine learning gain. Educational AI guided by clear teaching principles is more likely to produce learning that lasts. The Education Endowment Foundation is now funding research into cognitive offloading among 13- to 15-year-olds because the evidence for pupils is still developing. That caution is sensible. We should expect educational technology to demonstrate what children are learning, not merely what a system can produce for them.


    There is encouraging evidence when pedagogy leads the design. A 2025 randomised controlled trial in an undergraduate physics course found that students using an expert-designed AI tutor learned more in less time than students in an active-learning class. The researchers were careful about the limits of the study, but their central design lesson matters: teaching principles had been deliberately built into the tutor. The model was not simply asked to be helpful.


    What a teacher does when a child is stuck


    During more than 20 years in classrooms, I heard the words "I don't understand" thousands of times. My first response was rarely to repeat the same explanation more loudly or to give the answer. I asked questions. What do you already know? Show me the point where it stopped making sense. Why did you choose that answer? Can you explain what you think is happening?


    "I don't understand" is not a diagnosis. Two children can give exactly the same wrong answer for completely different reasons. One may have forgotten a fact. Another may be missing knowledge from a previous topic. A third may hold a confident misconception. Each child needs a different next step. We wrote about why diagnosis first learning changes everything in this article.


    Sometimes the quickest way forward in learning is to go backwards. A teacher might revisit a prerequisite idea, use a model or analogy, break the problem into smaller steps, create deliberate conflict with a misconception, provide guided practice or offer greater challenge. Only after locating the cause can we choose the right response.


    That level of diagnosis is difficult to provide continuously to 30 children at different starting points while moving through a large curriculum. Subject shortages add another layer. Recent reporting found that a quarter of state secondary schools in England lacked a dedicated physics teacher. That is not a criticism of teachers who are asked to teach outside their specialism. It is a constraint within the system, and children can feel the effect when an earlier gap remains hidden and later learning depends on it.


    Pedagogy came before the technology


    SideCog began with a teaching question. What does an expert teacher do when a learner is stuck, and how can technology make more of that individual attention available outside the classroom?


    Our team brings more than 50 years of combined classroom experience, including leadership, subject teaching and one-to-one tutoring. We did not start with a general chatbot and add education language around it. We started by defining the learning process the system must follow.

    1. Establish what the learner already knows and can do.

    2. Use questioning to locate a knowledge gap or misconception.

    3. Choose the next best action rather than serving more of the same.

    4. Reteach with an appropriate explanation, model, analogy or scaffold.

    5. Check understanding and ask the learner to retrieve, explain and apply the idea.

    6. Return to important knowledge through spacing and interleaving until it is secure.

    This is adaptive teaching in practice. It is more than changing the difficulty of the next question. The learning goal remains ambitious, while the route, explanation, level of support and timing respond to the child in front of us.


    The research has to change how the product behaves


    Daniel Willingham writes in Why Don't Students Like School that "memory is the residue of thought". The line is short, but its consequence for AI is enormous. If the machine does the thinking, what residue is left for the child? SideCog therefore has to ask questions that make the learner retrieve, compare, explain and connect ideas.  Further reading on how everything comes back to memory.


    Peter Brown, Henry Roediger and Mark McDaniel make a related point in Make It Stick: "Learning is deeper and more durable when it is effortful." That does not mean leaving a child confused or making work difficult for its own sake. It means preserving productive effort and giving enough support for the learner to succeed without removing the thinking that creates learning.


    Barak Rosenshine's Principles of Instruction gives us practical disciplines: review previous learning, present new material in small steps, ask many questions, check understanding, provide scaffolds and move towards independent practice. Bloom's Taxonomy reminds us that recall is a foundation rather than the finishing line. Learners also need to apply, analyse and explain. Work brought together by the Learning Scientists supports retrieval practice, spacing, interleaving, elaboration, concrete examples and dual coding.


    These are not references added to a product page after the software has been built. They determine how SideCog behaves. They shape the questions it asks, the order in which it teaches, when it offers support, how it revisits earlier learning and what it counts as evidence of understanding.


    Why a closed learning environment matters


    A general-purpose AI tool is designed to be broadly helpful. Ask for an answer and it will usually try to provide one. That is useful in many settings, but it is a poor default for learning. A good teacher sometimes helps by refusing to give the answer immediately. The child may need a question, a hint, a simpler example or a return to an earlier concept.


    SideCog uses AI inside a closed, curriculum-aligned environment. The AI is trained and constrained around a defined age range, selected curriculum content and teacher-designed learning strategies. It is not free to send a child wandering through the open internet or to invent its own educational purpose. Its role is to question, diagnose, teach, check and revisit within boundaries we have set.


    A walled garden does not make any AI infallible. It narrows the task, makes behaviour easier to test and gives us a clear standard against which to improve it. This approach also reflects the Department for Education's product safety standards, which ask educational AI providers to state their intended purpose, target age group and learning focus, and to address risks including cognitive development, inappropriate content and overreliance.


    The technology should adapt within pedagogical boundaries. It should not be allowed to decide what good teaching is. That judgement comes from teachers, learning science and continuing evidence from real learners.


    Supporting teachers and making learning visible


    SideCog is not designed to replace teachers. The Department for Education is right that AI cannot replace human judgement or deep subject knowledge. Nor can it reproduce the relationships, encouragement and understanding that a skilled teacher brings to a classroom.


    It can, however, extend the reach of good teaching. It can give a child more time to think in private, where a wrong answer feels safer. It can keep asking careful questions when a lesson has moved on. It can identify an older gap beneath a current difficulty and then teach from the right starting point. For a child taught by a non-specialist, it can provide consistent subject-specific explanation and practice alongside school, not in place of it.


    It can also give families something they often lack: visibility. Parents do not need another dashboard showing only minutes spent, questions completed or a percentage score. They need to understand what their child knows, where misconceptions remain, what has improved and what should happen next. That is why SideCog diagnoses before it teaches and why the parent view matters as much as the activity on the learner screen.


    The test for educational AI


    When I assess an AI learning tool, I want to know:

    • Did the AI do the thinking, or did it make the learner think?

    • Did it diagnose the cause of an error, or simply mark it wrong?

    • Did it connect new learning to the knowledge already in memory?

    • Did it adapt the support without lowering the ambition?

    • Could the learner retrieve and use the knowledge later without the AI?

    • Was the experience bounded, age-appropriate and transparent?

    If the answer to those questions is no, a fast and fluent response is not educational progress. It is task completion.


    The current scrutiny of AI in education is necessary. The risks of cognitive offloading, unsafe content, inaccurate answers and overdependence are real. We should not respond by pretending that every use of AI is the same. Open-ended answer generation and pedagogy-led educational AI are different propositions and should be judged differently.


    SideCog is built by teachers because education does not need an AI that simply knows how to answer. It needs a system that knows when to question, when to explain, when to step back and when to let the learner do the hard thinking.


    The real test is not what a child can produce while the AI is open. It is what they understand and can do after the screen closes.


    Sources and further reading

    OECD Digital Education Outlook 2026  Exploring Effective Uses of Generative AI in Education

    OECD PISA 2025 Results Volume I and the relationship between AI use and student performance

    The Verge  Students who use AI generally score worse at school  9 September 2026

    Education Endowment Foundation  Research into generative AI and learner cognition  8 June 2026

    Department for Education  Generative artificial intelligence in education

    Department for Education  Generative AI product safety standards

    Scientific Reports  AI tutoring outperforms in class active learning  2025

    Daniel T Willingham  Why Don't Students Like School  Second edition

    Peter C Brown  Henry L Roediger III and Mark A McDaniel  Make It Stick

    Barak Rosenshine  Principles of Instruction

    The Learning Scientists  Six Strategies for Effective Learning

    Department for Education  Subjects taught and specialist teachers and hours  2025 to 2026

    The Guardian  Quarter of schools in England lack a physics teacher  2 September 2025


    J

    Written by

    Julian Dance

    Rate this article
    Was this helpful?

    Sign in for free to react to this article.

    Browse all SideCog articles