Adaptive Learning Gets Real

The Evolution of EdTech — Part 10 of 13

By Clarence Stephen | May 2026

2,200 words | 9 min read

For seventy years, the promise has been the same: a machine that adapts to the learner. Notices confusion and adjusts. Recognizes mastery and advances. Infinite patience, infinite availability, meeting every student where they are.

Pressey imagined it in 1924. Skinner designed for it in 1954. Papert aspired to it in 1980. Khan Academy approximated it in 2008. For seventy years, the promise outran the technology.

In 2023, the technology caught up. Not to the full promise — that may never happen. But to a threshold where the gap between "AI tutor" and "no tutor at all" became large enough to matter for millions of students.

The Evidence Problem

For most of EdTech's history, the evidence base for adaptive learning was thin, contradictory, or nonexistent. Companies made extravagant claims. Peer-reviewed studies were scarce. When rigorous research did appear, results were modest — small effect sizes, narrow populations, short durations.

This wasn't because the systems were useless. It was because they were crude. Pre-AI adaptive learning was branching logic: if the student answers A incorrectly, show question B instead of C. These rule-based systems could handle straightforward hierarchies (multiplication before division) but couldn't navigate the messy, nonlinear reality of how humans actually learn.

Here's the thing. A student who fails a fractions problem may not lack fraction skills. They might lack confidence. They might be hungry. They might be confused by the wording, not the math. A branching algorithm can't tell the difference. A large language model — in principle — can. And that distinction is enormous, because the gap between a diagnostic conversation and a branching quiz is the gap between a tutor and a worksheet.

Khanmigo: The Proof of Concept

In March 2023, Sal Khan demonstrated Khanmigo — Khan Academy's AI tutor, built on GPT-4 — in a TED talk that went viral. The demo was striking. Khanmigo didn't give answers. It asked guiding questions. It noticed when students were stuck and offered hints instead of solutions. It adjusted explanations to the student's level.

In other words, it was performing the core skill of effective human tutoring. No previous educational software had pulled that off.

Khanmigo launched as a pilot in a handful of districts for the 2023-2024 school year. Initial results were promising — a study found significant learning gains across all conditions, though it didn't establish statistically significant differences between Khanmigo and Khan Academy's existing tools. Students did report finding the AI tutor helpful and engaging.

More telling: the adoption trajectory. Usage jumped from 40,000 to 700,000 K-12 students in the 2024-2025 school year, expanding to hundreds of districts and classrooms in India, Brazil, and the Philippines.

The approach was sound — guided questioning, step-by-step problem-solving, the Socratic method running on the newest technology. A genuine departure from the drill-and-practice model that had dominated EdTech for decades.

The Research Catches Up

The research matured fast in 2023-2024.

A systematic review of AI-driven tutoring systems in K-12 found overall effects "generally positive." A 2024 study with 300 high school students showed significant improvements in problem-solving, critical thinking, and logical reasoning.

I want to be honest about what this means. "Generally positive" isn't "transformative." Effect sizes were real but comparable to other evidence-based interventions — smaller than high-quality human tutoring, larger than most previous EdTech tools. Studies were short-term, populations limited, long-term effects unknown.

But for a field that had been making unsupported claims for decades, the arrival of actual evidence — even modest evidence — was an inflection point.

MagicSchool and the Teacher Tools

While AI tutoring grabbed headlines, a parallel development may prove equally important: AI tools designed for teachers.

MagicSchool AI launched with over 80 tools for teaching tasks — lesson plans, assessments, rubric templates, text leveling, report card comments. Not glamorous. Practical. The digital equivalent of a teaching assistant who handles the paperwork so the teacher can actually teach.

The Text Leveler is worth calling out. It takes any passage and rewrites it at different grade levels, letting teachers differentiate reading assignments for students of varying abilities. That task used to take hours of manual work. MagicSchool does it in seconds.

American teachers work an average of 54 hours per week, and a huge chunk of that is planning, grading, and admin — not instruction. AI tools that automate even part of that non-teaching labor could reshape the profession. Not by replacing teachers, but by giving them back time for what only humans can do: build relationships, spark curiosity, respond to the kid who's clearly having a bad day.

What Personalization Actually Means

"Personalization" has been so abused by the EdTech industry that the word barely means anything anymore. For most of the field's history, "personalized learning" meant choosing from a menu of pre-built options, or getting routed through branching algorithms based on right/wrong answers.

AI-powered adaptive learning offers something different: the ability to respond to how a student is thinking, not just what they answer.

When a student tells a human tutor "I don't understand fractions," the tutor doesn't repeat the explanation louder. They probe: what part is confusing? The concept that a fraction is part of a whole? The notation? The relationship between numerator and denominator? They diagnose the specific misconception and address it directly.

Large language models can approximate that conversation. Not perfectly — they lack the social intelligence and emotional awareness of a great human tutor. But they can do it at 2:00 AM, for millions of students simultaneously, at a marginal cost approaching zero.

That's what makes this moment different from every previous wave of EdTech hype. The technology has crossed a threshold where personalized, conversational, adaptive instruction is possible at scale. Not perfect. Not human-equivalent. But meaningfully better than what most students get when they're stuck on a problem at home with no one to ask.

The Benjamin Bloom Question

In 1984, Benjamin Bloom published a finding that has haunted education policy ever since. Students receiving one-on-one tutoring from a skilled human performed two standard deviations above students in a typical classroom. The average tutored student outperformed 98 percent of the control group.

Bloom called this the "2 Sigma Problem": finding group instruction methods as effective as individual tutoring. Forty years later, no one has solved it.

AI tutoring won't solve it either. Current evidence shows effect sizes of 0.2 to 0.5 standard deviations — real but far short of 2 sigma. The relationship, the emotional attunement, the ability to read frustration or excitement — these aspects of human tutoring may represent a permanent gap.

But here's where the framing matters. In practice, almost no students get one-on-one human tutoring. It's too expensive, too labor-intensive, too dependent on tutor quality. The relevant comparison isn't "AI tutor versus human tutor." It's "AI tutor versus no tutor at all."

For the millions of students who go home to empty apartments, whose parents work multiple jobs, who attend underfunded schools with overwhelmed teachers — an AI that can explain fractions at midnight with infinite patience isn't a compromise. It's a breakthrough.

The Remaining Gaps

The adaptive learning revolution is real. It also has gaps that I'm not going to gloss over.

The equity gap persists. AI tutoring requires devices and internet access — the same resources that the students most in need of tutoring often lack. The technology can't serve students it can't reach.

Quality varies wildly. For every Khanmigo, there are dozens of poorly designed, inadequately tested tools optimized for engagement metrics instead of learning outcomes. The market is noisy.

The content is narrow. Most AI tutoring focuses on STEM subjects where problems have clear answers. Humanities, creative writing, critical analysis, social-emotional learning — all harder to adapt, all less well-served.

Reading remains unsolved. This is the gap that keeps me up at night. Despite all advances in AI tutoring, the American reading crisis continues to worsen. Adaptive math tools are plentiful. Adaptive reading tools — systems that help struggling readers engage with text, visualize content, and build comprehension — are nearly nonexistent. The 130 million American adults with limited literacy skills aren't being served by Khanmigo or any comparable tool.

The reasons are technical and philosophical. Math has clear structure: concepts build predictably, problems have definitive answers. Reading is contextual, subjective, personal. What makes a text difficult for one reader may have nothing to do with what makes it difficult for another. The technology for adaptive reading intervention exists in theory. In practice, almost no one has built it.

Almost.

Next in the series: "The Reading Crisis Nobody's Solving" — the numbers are staggering, the consequences are devastating, and the EdTech industry is looking the other way.

Sources

Clarence Stephen is the founder of Readify, an AI-powered literary platform. Yale B.Sc. Physics. Previously: Tiger Global, Morgan Stanley. Learn more at ireadifybooks.com

What could AI do for your classroom?

Readify uses local AI to adapt reading material to each student's level — generating illustrations, simplifying vocabulary, and tracking comprehension in real time. No student data leaves the building.

Bring Readify to Your School Try the Platform

Liked this article?

Get the full Evolution of EdTech series delivered to your inbox, plus weekly insights on AI in education.

Subscribe to Readify Branches