AI Tutors Support 16 Percent of Learning. What About the Other 84 Percent?
A parent asked me recently whether they should sign their child up for an AI tutoring service. The marketing was persuasive: personalized learning pathways, instant feedback, mastery-based progression. It sounded like the future of education had arrived. My answer surprised them. I said that AI tutoring could be genuinely useful for part of their child’s learning, but that it would only ever address a fraction of what it means to become an intelligent, capable human being. The question we should all be asking is not whether AI can tutor, but what exactly it is tutoring, and what it is leaving out.
The excitement around AI tutors has a long history. Benjamin Bloom’s 1984 “2 sigma” finding, which suggested that one-to-one tutoring could lift the average student two standard deviations above conventionally taught peers, is routinely invoked to justify AI-powered personalization. Yet that finding deserves more scrutiny than it typically receives. The original study involved graduate students tutoring small groups under highly specific conditions, using mastery learning techniques alongside the tutoring itself. Subsequent attempts to replicate the full effect have produced mixed results, and the finding is better understood as an aspirational benchmark than a reliable empirical standard. What it does tell us is that individualized attention matters. The question is what kind of attention, and directed towards what.
As Tom Chatfield argued in a recent Sage white paper on AI and the future of pedagogy, we need to begin with what works in education before asking what technology can contribute. The human fundamentals of learning, including metacognition, social sense-making, and cognitive load management, should drive our approach to technology rather than the other way around.
Current AI tutors can generate explanations, provide hints, and offer feedback on student work. In carefully constrained settings, AI-generated help can produce learning gains comparable to human-authored help for particular skills. For delivering academic content, practising procedures, and drilling factual knowledge, AI tutors can be effective. But this is not nearly enough.
Academic knowledge represents only one element of what it means to be intelligent. I have argued, drawing on decades of research in cognitive science, developmental psychology, and education, that human intelligence is best understood as an interwoven model comprising seven elements. The first, knowledge about the world, is the domain where AI tutors operate. The remaining six are where things become far more consequential: our personal epistemology (understanding what knowledge is and how it is constructed), our social intelligence (collaborating, debating, and learning from others), our metacognitive awareness (knowing what we know and do not know), our metasubjective intelligence (recognizing and regulating our emotions and motivation), our metacontextual awareness (understanding how our environment shapes our thinking), and our accurate perceived self-efficacy (making sound judgements about our own capabilities). Together, these six elements constitute the vast majority of what makes human intelligence distinctive.
There is another way to appreciate just how narrow AI tutoring is: by examining not the kinds of intelligence we wish to develop, but the acts of learning through which development happens. The CAPITAL framework, developed by Andrew Manches and colleagues, identified 19 distinct acts of learning across educational settings. These range from personal acts such as browsing, annotation, rehearsal, construction, and reflection, through dialogic acts such as exposition and tutorial interaction, to social acts including performance, participation in communities of practice, and genuine collaboration, and scenarioed acts such as cross-contextual transfer, simulation, and inquiry.
When we map AI tutoring against this landscape, its reach is strikingly limited. AI tutors operate primarily within rehearsal, exposition, and partially within tutorial interaction and assessment. They cannot support the social acts of learning through which learners construct knowledge together. They cannot facilitate the scenarioed acts that require transfer across real-world contexts. And they address only a narrow band of the personal acts, largely missing reflection, construction, and the playful, exploratory engagement that characterises deep learning.
An AI tutor cannot model what it means to wrestle with the provisional nature of knowledge. It cannot teach a student to debate conflicting evidence, because it does not understand what evidence is. It cannot develop the emotional regulation needed to persist through genuine intellectual difficulty, because it has no emotions and no experience of difficulty. These are not temporary limitations. They reflect fundamental differences between artificial and human intelligence.
The risk is that enthusiasm for AI concentrates investment on precisely the element of intelligence that machines already handle well, while neglecting the elements that only human educators can develop. We may produce students who recall facts efficiently but lack the metacognitive, social, and epistemic capacities they need in a world increasingly shaped by AI. We will be training humans to compete with machines on the machines’ own terms.
The evidence suggests this is a real danger. Recent research has found that AI-assisted learners demonstrate reduced self-monitoring, increased procrastination, and what has been described as “metacognitive laziness.” When the AI does the cognitive heavy lifting, learners complete tasks faster without building underlying competence. Remove the AI, and the gains evaporate.
So what should we do? First, welcome AI tutors for what they genuinely do well: delivering personalised academic content and providing timely feedback on structured tasks. These contributions can free human educators to focus on what they do best.
Second, invest with urgency in developing the pedagogical approaches and assessment tools needed to cultivate the other six elements of human intelligence and the full range of learning acts that make education transformative. We need teachers equipped to develop their students’ metacognition, their capacity for collaborative inquiry, their emotional resilience, and their ability to judge their own learning accurately. AI can generate data that helps teachers do this more effectively. But the teaching itself must remain profoundly human.
Third, reform assessment systems. As long as examinations reward the recall and reproduction of knowledge, schools will focus on what AI tutors deliver. If we want education to develop the full breadth of human intelligence, we need assessments that value metacognition, collaboration, epistemological sophistication, and self-regulation.
The promise of AI in education is genuine, but far more limited than the hype suggests. AI tutors address perhaps 16 percent of what it means to develop an intelligent human being, and a handful of the 19 acts through which real learning happens. The remaining 84 percent requires investment in human educators, in the science of learning, and in assessment systems that recognize the full complexity of human intelligence. If we get this balance right, AI could be the catalyst that finally compels education to value what makes us distinctively human. If we get it wrong, we will have built the most sophisticated system ever devised for teaching people the things that matter least.
