What an adaptive AI tutor adapts to

Three things a tutor can adapt

A tutor is adaptive when what it does next depends on what you just did. That can happen at three sizes:

• Within a problem. It reads each step, not only the final answer, and responds to the first step that goes wrong.
• Across problems. It gives more or less help, or picks the next problem, based on how the last ones went.
• Across sessions. It remembers the mistakes you've made before and brings them back when they matter.

alai is an iPad and web notes app, and its tutor, Ally, adapts at all three sizes. The research below explains why each one matters, and the last sections list what Ally keeps and how to clear it.

Feedback on each step has the strongest evidence

Computer tutors that follow a student's steps through a problem are called intelligent tutoring systems, and they've been studied for decades. Two large reviews found clear gains:

• VanLehn (2011) reviewed experiments comparing human tutors, computer tutors and no tutoring. Human tutoring raised test scores by an effect size of 0.79. Intelligent tutoring systems, which mostly respond step by step, raised them by 0.76, so they were nearly as effective as a human tutor.
• Kulik and Fletcher (2016) pooled 50 controlled evaluations. The median effect was 0.66 standard deviations, enough to move a typical student from the 50th to the 75th percentile. The size of the gain depended a lot on whether it was measured with a test written for the study or with a standardized test.

Tutors built on large language models are now being tested the same way. In a 2025 randomized trial at Harvard, 194 students in an introductory physics course learned two topics either from an AI tutor at home or in an active-learning class. The median learning gain with the AI tutor was more than twice as large, and students got there in less time. The tutor was built for the course, with prompts its instructors wrote, and the authors say the result may not hold in every setting.

The design matters as much as the AI. In a 2025 PNAS study, high school students who practiced math with a ChatGPT-style chat scored 17% lower on a later exam than students with no AI, while students whose tutor was set up to give hints scored about the same as the no-AI group. A tutor that adapts by doing more of the work for you isn't helping you learn.

Adjusting how much help you get

A good tutor gives the smallest help that gets you moving, and more only if you're still stuck. Ally works this way while you're solving your own problem:

1. Ally reads your page after you pause, finds the first step that doesn't follow from the one before, highlights it and asks a short question about it.
2. If you're still stuck, the next hint is smaller and more concrete, and Ally never says the same line twice.
3. If you ask out loud with Hold to talk and a spoken hint doesn't land, Ally steps up to a fill-in-the-blank outline of your own problem, then to a worked example of a different problem.

When a step is right, Ally stays quiet. Ally is set up not to say the corrected step or the final answer while you work, though like any AI, Ally can make mistakes. The guide to homework help that gives hints instead of answers explains why.

Adapting to the kind of help that works for you

Some students get unstuck when someone points at the page. Others need the idea talked through, a guiding question or a worked example. Ally tries each of these four ways of helping on your first problems, then keeps using the one after which you most often get problems right on the first try, with speed as the tiebreaker. A way of helping that had one bad session gets tried again now and then, so a single day doesn't decide it for good.

That's different from teaching to a learning style. A 2008 review by Pashler and colleagues found no adequate evidence that matching lessons to a student's preferred style, such as being a "visual learner", improves learning. Ally doesn't label you as a type of learner. Ally keeps a record of what helped on your own problems and keeps checking it.

Remembering your mistakes between sessions

When you're signed in, alai keeps a learning history made of short codes: the topic, the mistakes you made, that a hint was shown and what fixed it. It never holds your handwriting, transcripts or images, as the privacy policy explains. Ally uses it in three ways:

• The greeting. A returning session opens with the part that was tricky last time, so you can start there.
• Repeat slips. When you make a mistake you've made before, Ally can say so. For example: "Like last time with distributing, does the 5 in 5(x + 2) reach the 2 too?"
• Practice. On iPad, Ally can write two or three new problems for a recent slip, each built so the same slip is the trap.

Coming back to old mistakes later fits what's known about studying. Dunlosky and colleagues (2013) rated ten common study techniques, and only two earned their highest rating: practice testing and spreading practice out over time. Working a new problem on a slip from an earlier session is both.

What you can see and change

• The Memory drawer in each session shows what Ally is using.
• What Ally knows lists facts such as your classes, and you can edit or delete each one.
• Reset Ally's memory clears the learning history on every device. Your notes stay.

What adaptive tutors can't do yet

• They can misjudge. Ally can misread messy handwriting or judge a step wrongly, and a wrong call can end up in the history. That's one reason you can reset it.
• They need a few sessions. The way of helping settles only after Ally has some history, so the first sessions are partly trying things out.
• Gains depend on the test. As Kulik and Fletcher found, tutors look better on tests that match what they taught than on broad standardized tests.
• alai has no outcome data yet. alai hasn't published measurements of how much students learn with Ally. The studies above are about tutoring in general, not about alai.