Science of Learning Guide Bridges GenAI Gap for Secondary Teachers

At the National Education Summit in Melbourne, Professor Miriam Tanti sat down to discuss a problem that has been quietly building in Australian classrooms: teachers and students are using generative AI, but not together, and not in ways that align with what we know about how learning actually works.

Tanti, Deputy Dean of the School of Education at La Trobe University and co-director of the university's AI Institute, has just released The Science of Learning and Generative AI, a free guide designed to bridge that gap. Her path to this work is instructive. She began as a high school computing teacher in the early 2000s, when the biggest classroom management challenge was students stealing mouse balls from computer lab equipment. She was in schools in 2008 when Prime Minister Kevin Rudd announced the Digital Education Revolution—handing every student a laptop without the professional learning, infrastructure planning, or pedagogical thinking to make it work. That experience, she says, drove her toward academia and ultimately toward a PhD focused on what she calls a "slow pedagogical approach" to technology in education: deliberate, reflective, human-centered, built on time and connection rather than speed and novelty.

The rationale for the new guide emerged from ground-level conversations with teachers and school systems. Teachers were asking for evidence-informed practice—not broad policy statements about responsible AI use, but concrete answers: What does this actually look like in my year 10 classroom? What does the research tell me works? What Tanti and her colleagues discovered was a troubling disconnect. In one education system where Microsoft Copilot was deployed to all teachers, usage was high and sustained, but it was focused almost entirely on administrative efficiency: drafting emails, collating documents, differentiating lessons. Teachers were not bringing generative AI into their teaching. Meanwhile, students in years 8 through 12 were using AI tools on their own—converting notes into podcasts, creating flashcards for revision, asking the tool to explain concepts they didn't understand in class—all through self-teaching and trial-and-error, with no formal instruction. The guide attempts to close that gap by anchoring everything in cognitive science rather than technology features.

The framework at the heart of the guide works outward from the architecture of learning itself. Attention acts as a filter, pulling forward what seems most relevant. Whatever passes through that filter enters working memory, a small and temporary space where thinking happens. For information to stick, students must do something with it—connect it to what they already know, rehearse it, apply it, answer questions about it. That active processing moves information into long-term memory, where it grafts onto existing knowledge structures called schemas. Each time this happens, the schema deepens and expands, freeing up working memory to tackle harder problems and think more critically. This is the foundation. The risk with generative AI, Tanti explains, is that students will outsource these cognitive processes to the tool. Because ChatGPT can produce a fluent, comprehensive answer in seconds, the temptation to offload is high. When students completely bypass the cognitive work—what researchers call detrimental cognitive offloading—they lose the desirable difficulty that makes learning stick. Equally dangerous is the illusion of mastery: a student reads a fluent AI-generated answer, feels confident they understand it, but their working memory has done no processing. There is no schema building, no durability of knowledge, only false confidence.

The guide identifies four key areas where generative AI intersects with learning: attention and cognitive load, memory and knowledge, knowledge transfer, and metacognition—the ability to notice your own thinking and act on it. Metacognition is the thread that holds the others together. Without it, even a well-designed task with managed cognitive load will fail if a student doesn't recognize they've stopped understanding or doesn't retrieve knowledge from memory. Research cited in the guide is sobering: students with limited domain knowledge tend to have weaker metacognitive habits to begin with, making them more likely to engage in detrimental offloading. A 2024 study by Fan and colleagues gave 117 university students the same writing task under different conditions—some used ChatGPT, some worked with a human expert, some used a writing analytics tool, some had no support. The ChatGPT group produced the best essays. But when the technology was removed, they couldn't transfer that knowledge to other settings. The process data revealed why: students had handed over the monitoring and evaluation work to the tool. They never noticed when things weren't working or decided what to fix next. The researchers called it "metacognitive laziness."

Tanti illustrates how intentional use looks different. In a year 10 science class teaching greenhouse gases for the first time, a teacher also introduced ChatGPT for the first time. Because both were new, both required explicit teaching. The lesson began with no AI—activating what students already knew about greenhouse gases, priming their long-term memory to receive new information. Then the teacher modeled expert use: not just giving a prompt, but explaining why that prompt was crafted that way, showing what the chatbot returned, asking follow-up questions aloud, demonstrating how to use the tool for feedback and revision, never for answers. Throughout the lesson, students moved in and out of AI-free moments—low-stakes retrieval practice, turning to a partner to explain what they'd learned. The teacher managed cognitive load by removing extraneous information, then gradually released responsibility as students gained confidence. The focus never shifted from learning; generative AI was always in service to it, never the center.

Tanti's team is pursuing two significant research directions. They are developing an instructional hierarchy for AI thinking—not necessarily the tools themselves, but the reasoning needed to engage with evolving technology—that spans from primary school through high school. More immediately, they are working with cognitive neuroscientists and year 9 students, using EEG and eye-tracking software to measure what actually happens in the brain under three conditions: completing a literacy or numeracy test with no AI, with completely detrimental offloading to ChatGPT, and with scaffolded intentional use. The results, she says, are promising and affirming the work already underway. That data is expected within months. For teachers watching generative AI reshape their classrooms without clear guidance, that evidence cannot come soon enough.