Guided Learning and AI: What the Evidence Says About How Students Learn Best
By Julian Dance —
Tags: EdTech, Student Progress, Cognitive Science, AI in Education, Hyperpersonalisation, Pedagogy, Personalised Learning
As AI becomes increasingly common in education, an important question remains: what actually helps students learn? This article explores why guided learning, scaffolding and active thinking matter more than simply giving learners answers.
By Julian Dance, Co-Founder of SideCog, Head of Science and Head of Big Ideas Guided Learning - Why Pedagogy Matters More Than Technology Artificial intelligence is rapidly becoming part of education. New tools appear almost daily, each promising to personalise learning, answer questions instantly, and help students make faster progress. The technology is impressive. But there is an important question that sits beneath the excitement: What actually helps people learn? For decades, educational research has shown that learning is not simply about receiving information. Understanding develops when learners think hard about ideas, connect them to prior knowledge, retrieve information from memory, identify misconceptions and apply what they know in new contexts. The challenge for educational technology is not whether it can provide answers. It is whether it can support these processes. This is why guided learning matters. The Problem with Answer Machines One of the most common concerns about generative AI in education is that it can make learning appear easier without necessarily making learners more knowledgeable. If a student asks a question and immediately receives the solution, they may complete the task, but they may not develop the understanding required to solve similar problems independently in the future. Educational psychologists often refer to this as the distinction between performance and learning. A learner may perform well in the moment because support is available, yet retain very little once that support is removed. This concern is increasingly being recognised by researchers studying AI in education. A recent study conducted by Google DeepMind, Fab AI and the Sierra Leone Ministry of Education specifically examined whether AI could improve learning outcomes without encouraging students to bypass the thinking required for learning. The researchers designed a "Guided Learning" approach intended to promote understanding rather than simply provide answers. What Is Guided Learning? Guided learning is not a new idea. Its roots can be found in decades of research into scaffolding, cognitive apprenticeship, formative assessment and instructional dialogue. Rather than giving students answers immediately, guided learning supports them through carefully structured questioning, prompts and feedback. The learner remains responsible for the cognitive work. The role of the teacher - or increasingly, the technology - is to guide that thinking process. This approach aligns closely with what educational research tells us about effective learning: Learning requires active cognitive engagement. Prior knowledge strongly influences new learning. Misconceptions must be identified and addressed. Feedback is most effective when it helps learners think rather than simply reveals answers. Retrieval and application strengthen long-term memory. These principles have been consistently supported across learning science and cognitive psychology research. Evidence from Sierra Leone The recent randomised controlled trial conducted in Sierra Leone provides one of the strongest pieces of evidence currently available for AI-supported guided learning. The study involved 1,763 junior secondary students across 12 schools over an eight-week period. Students using Gemini's Guided Learning feature were compared with a control group following normal classroom activities. The intervention was teacher-led, with educators setting objectives, designing lessons and facilitating learning. AI was used as a support tool rather than a replacement for teaching. The findings were notable. Students using Guided Learning achieved gains of approximately 0.258 standard deviations in mathematics performance, equivalent to around 1.2 to 1.7 years of typical learning progress during the eight-week trial. Just as importantly, the interaction data suggested that students were not primarily using the system to obtain answers. Researchers analysed more than 113,000 interactions and found that: Students were building conceptual understanding in 91.4% of conversations. The AI responded with scaffolding questions in 76% of messages. Direct solutions were provided in only 2% of interactions. These findings matter because they suggest that the learning gains were associated with learners engaging in the thinking process rather than outsourcing it. The Importance of Scaffolding One feature repeatedly highlighted in both educational research and recent AI studies is scaffolding. Scaffolding refers to temporary support that helps learners achieve something they could not yet do independently. As understanding develops, that support is gradually removed. Effective teachers do this constantly. They ask questions rather than provide answers. They break complex tasks into manageable steps. They identify misconceptions before they become embedded. The Sierra Leone study deliberately incorporated these principles, using Socratic questioning to maintain student ownership of the learning process. Similar findings have emerged from research conducted by Eedi and Google DeepMind in UK schools. In a 2025 exploratory randomised controlled trial, students supported by a pedagogy-tuned AI model performed at least as well as students receiving support from human tutors alone. Researchers reported that the model was particularly effective at generating Socratic questions that encouraged deeper reflection and reasoning. Why Curriculum Structure Matters Guided learning is most effective when it is anchored to a coherent curriculum. Learning is not a collection of isolated facts. Knowledge builds over time. Understanding photosynthesis depends upon understanding cells. Understanding algebra depends upon prior numerical reasoning. Understanding forces depends upon prior concepts of motion and energy. This is why curriculum architecture matters. To identify misconceptions accurately, a system must understand what should have been learned previously, what concepts are connected, and what knowledge is required next. Without this structure, personalisation becomes little more than content recommendation. With it, learning can become genuinely adaptive. The focus shifts from asking, "What would you like to learn?" to asking, "What understanding is missing and what should happen next?" Technology Should Support Learning Science The most interesting finding emerging from current research is not that AI can answer questions. We already knew that. The more important finding is that AI appears to be most effective when it applies established pedagogical principles rather than attempting to replace them. The strongest results are coming from systems that: Encourage active thinking. Use questioning instead of answer-giving. Identify misconceptions. Scaffold understanding. Support retrieval and application. Work alongside teachers rather than replacing them. In other words, the technology succeeds when it behaves more like an effective teacher. Looking Ahead The conversation around AI in education often focuses on models, features and technological capability. Those things matter. But the evidence increasingly suggests that pedagogy matters more. The question is not whether AI can teach. The question is whether it can support the conditions under which learning occurs. The emerging research from Sierra Leone, the United Kingdom and elsewhere suggests that guided learning may be one of the most promising approaches currently available. However, researchers continue to emphasise the need for further high-quality trials, long-term studies and careful evaluation of outcomes. For educators, parents and developers alike, the message is clear: Technology should not do the thinking for learners. It should help learners do the thinking for themselves.