AI in the classroom

The oldest problem in education is a staffing problem

The most reliable finding in instructional research is also the least surprising: students learn dramatically more with a capable tutor at their side than in a lecture hall. The interesting finding came later. VanLehn's 2011 review in Educational Psychologistshowed that intelligent tutoring systems approach the effectiveness of one-on-one human tutoring, but only when their moves are pedagogically principled rather than merely reactive. Kulik and Fletcher reached a similar conclusion in their 2016 meta-analysis. The lineage runs back to Koedinger and Anderson's Cognitive Tutor work in the 1990s, which put research-grade adaptivity in front of real algebra classrooms and measured what happened.

For thirty years, that kind of adaptivity was available only to schools inside funded studies. The systems were expensive to build because every adaptive move had to be hand-engineered by a team of cognitive scientists. What has changed is not the pedagogy. The pedagogy was settled. What changed is that the engineering cost of a principled adaptive move collapsed. That is the opportunity, stated plainly: the instructional pattern that used to require a lab can now ship inside an ordinary lesson, in an ordinary school, drafted by an ordinary teacher on a Tuesday night.

Adaptivity does not mean branching flowcharts

A persistent belief holds that adaptive lessons require the teacher to author a decision tree: if question three is wrong, show page 4a; if the student picked C, show 4b. Anyone who has tried this knows how it ends. The tree explodes combinatorially, and only full-time curriculum teams can afford to tend it.

The research suggests a different shape. Instead of a tree, a small set of response patterns: insert a worked example after a failed first attempt, offer a single hint when a student idles past the thinking window, route back to the relevant teaching page when the same misconception appears twice, pose one transfer-level question when a student finishes suspiciously fast. Learning scientists call this just-in-time scaffolding. The help arrives at the moment of need, not before, which wastes attention, and not after, which lets a misconception consolidate. In Prism these patterns are called triggers, and no teacher authors them. They are built into every lesson page and stay invisible unless a particular student needs one.

Help that stays inside the lesson

The second persistent belief is that classroom AI means a chatbot: a separate window students wander into and, too often, do not wander back from. The evidence on interruption is not kind to that model. Cognitive load theory, from Sweller and colleagues, predicts that forcing a reader to leave the text to get help taxes exactly the working memory the text was occupying. Conversely, Roy and Chi found that brief, prompted explanations delivered during reading improve both comprehension and retention.

The design conclusion is that help should surface inside the page, scoped to what the student is actually reading. A student who highlights a term gets a one-sentence definition informed by the surrounding lesson, so that "limit" means one thing in calculus and another on the track team. The response is deliberately short, one highlight at a time, and the tools refuse to operate on graded questions, which closes the obvious loophole. This is scaffolding in Pea's original sense: temporary support a learner sheds once it is no longer needed.

Measurement that cannot be gamed by clicking

Skeptics of learning analytics usually have a specific failure in mind: the dashboard number that is really just engagement (did the student click the buttons?) or really just correctness (a lucky multiple-choice guess looks identical to mastery). Both are partial, and both are gameable. The skepticism is deserved.

The answer from the research community is neither metric alone but a model. Bayesian Knowledge Tracing, introduced by Corbett and Anderson in 1995, treats what a student knows as a hidden state and updates the probability with every answer, accounting for slips and guesses. It has three decades of classroom validation behind it, and, as Khajah, Lindsey, and Mozer showed in 2016, it outperforms far fancier deep-learning models on the small datasets a real classroom actually produces. Prism's P-Score builds on that foundation and weights mastery alongside transfer, resilience, and engagement, the dimensions Fredricks and colleagues identified as the structure of genuine engagement. Just as important is what the score refuses to do: with two pieces of evidence it presents itself as tentative, and with none it does not appear at all. The same evidence rolls mastery bands up against course, district, and state standards, computed fresh at every level rather than averaged, so a confident score is never diluted by a thin one.

What AI drafting actually is

The loudest fear is about generation itself: that AI-drafted lessons are plausible-sounding slop, and that adopting them outsources teaching to a language model. The fear is reasonable if you imagine the AI writing freeform prose. Prism does not work that way, and the difference is the entire point.

A generated lesson is a sequence of typed blocks drawn from a finite pedagogical vocabulary: 24 block types, each a distinct instructional move with research behind it, from worked examples to retrieval practice to structured reflection, the moves Hattie and Donoghue's meta-analysis ranks among the highest-leverage available. Generation is constrained structurally: every lesson must include framing, instruction, and practice; every tagged standard must be covered by at least one graded block; no block type may repeat more than four times; and quiz content must be answerable from the source material the teacher supplied, because the model may rephrase a source but may not invent facts that are not in it. Every generated block is labeled, and every one waits in a draft the teacher can edit or delete before a student ever sees it. This is the arrangement Mollick and Mollick describe as AI augmenting the teacher's design role rather than replacing it.

The teacher is the point

Every mechanism described above shares one property: it redirects machine effort toward the parts of teaching that were never the good parts. Nobody became a teacher to hand-build branching trees, to grade the same misconception forty times, or to discover in June what a weekly mastery view would have shown in February. The promise of classroom AI, honestly stated, is not a robot teacher. It is a classroom where the adaptive moves fire on time, the evidence accumulates into something a parent can understand, and the teacher spends their hours on judgment, relationships, and the craft of explanation. The research has said for decades that this is what moves learning. The tooling has finally caught up.

For how all of this runs without surveilling students, see AI and privacy.