AI Tutors & Personalized Learning
For forty years we have known one-to-one tutoring works almost magically well, and we could never afford it. Large language models change the math - but only if we remember that the model was never the teacher.
The two-sigma ghost
In 1984 Benjamin Bloom published a number that has haunted education ever since. Students who learned with a personal tutor, working each topic to mastery before moving on, outperformed their classroom peers by about two standard deviations. Two sigma is not a tweak; it lifts a middling student to near the top of the class. Bloom turned the result into a dare to the whole field: find a way to teach groups as well as a tutor teaches one. We never did. The reason was always economic, not pedagogical - a tutor for every learner is a luxury no system can buy at scale.
So we built substitutes. Intelligent tutoring systems, beginning in the 1970s, were the most serious attempt: software that imitated a tutor, gave instant feedback, and adapted to the learner. Some genuinely worked. But each one was hand-carved for a single narrow subject, every concept and misconception and hint authored by experts over years. A system that mastered algebra knew nothing of poetry. The intelligence was real and the cost was ruinous, which is exactly why, after decades, intelligent tutors reached almost no one.
Why the math suddenly changed
Large language models broke the cost curve that doomed the old systems. A single model can hold a fluent conversation about almost anything, in many languages, shifting register from baffled beginner to impatient expert. The years of hand-authoring largely evaporate. You no longer build a tutor for one subject; you instruct a general model to tutor whatever you point it at. The scarce resource Bloom identified - a patient, knowledgeable interlocutor on demand - just became close to free.
That is a genuinely large opportunity, the most plausible answer to the two-sigma dare in forty years. It deserves to be stated plainly before the caveats, because the caveats are where most of the work lives, and it is easy to let them swallow the promise whole.
The model is not the teacher
Here is the trap. An LLM is fluent, and fluency reads like competence. Ask it for the derivative and it gives you the answer, fast and correct-sounding - which is, for a learner, often the worst possible response. The student learns only that the machine is quicker than they are. Worse, the model will state false things with serene confidence, and a learner is by definition unable to catch it; that is why they are learning. Left to its defaults, an LLM is an answer-dispenser with an occasional fabrication problem. That is not a tutor. It is the opposite of one.
The thing that turns the model into a tutor is the pedagogy you wrap around it, and that pedagogy is old and well understood. It is Socratic guidance - answering a question with a smaller question that moves the learner one step forward. It is scaffolding - giving just enough support to let someone do what they could not do alone, then removing it as they grow. It is formative checking - "in your own words, why does that work?" - before advancing. It is grounding the tutor in vetted course material so it teaches your syllabus and invents less. None of this lives in the model. All of it lives in the instructions, the persona, and the design you build on top. The difference between a chatbot and a tutor is entirely the craft you add.
Personalization that means something
The same craft, scaled, is what makes "personalized learning" more than a slogan. Real personalization is not friendlier wording; it is adapting the path and the pace - skipping what a learner already knows, lingering where they are shaky, never dragging anyone forward before they are ready and never holding them back once they are. That is mastery learning, made practical for many learners at once. It requires the tutor to remember - a learner model that accumulates what each person has grasped and stumbled over, the way a human tutor builds a picture of a student across a term. And it produces a quietly valuable by-product: a real-time map of where a whole cohort breaks down, telling a teacher what to reteach rather than merely what to re-grade.
The line we must not forget
For every gift, a corresponding duty. A tutor that hands over answers erodes the productive struggle that learning is made of, leaving students fluent-feeling and dependent. A confidently wrong explanation plants an error that can take years to undo. The "great equalizer" widens the gap it promised to close if access is unequal. And the learner model - that intimate record of a child's confusions - demands a standard of privacy closer to a school's than a product's. Each risk has a discipline that answers it: design for guidance not answers, ground the tutor and teach learners to verify, treat access as a requirement, protect the data as if it were your own child's, and above all measure whether learners actually learn more rather than merely use the tool more.
That last discipline points at the deepest truth here. The tutor handles what scales - explanation, drill, the question at eleven at night. The human handles what never should scale - motivation, judgement, the messy reasons a learner has stopped trying, the decision about what is worth learning at all. AI does not replace the teacher; it relocates the teacher to exactly the work machines are worst at. Bloom asked for group instruction as good as one-to-one tutoring. We may finally be able to deliver something close - but only if we keep remembering that the model was never the one doing the teaching.
No comments:
Post a Comment