How does AI personalize learning? A plain-English walkthrough

Last updated: July 13, 2026.

"AI personalizes your learning" has become a claim every vendor makes and almost none explain. Here is the plain version, no hand-waving: three real inputs, combined by a recommender, to pick one specific lesson for one specific person today, instead of assigning everyone the same playlist. If a product cannot explain which of these three it actually uses, the personalization claim is probably decoration.

The three inputs, honestly

Input 1: role and goals. What you do, and what you are trying to get better at. This is usually set during onboarding and refined as you go, not a one-time survey that goes stale. A recruiter and a finance analyst should not see the same "AI skills" lesson on day one, even if both selected "improve with AI" as a goal, because what that goal means in practice is different for each of them.

Input 2: demonstrated mastery, not a self-rating. Self-ratings are notoriously unreliable; most people rate themselves as above average at almost everything. A system that actually personalizes tracks what you have shown you know, skill by skill, from how you perform on real checks and exercises, and updates that estimate gradually as new evidence comes in.

Input 3: what you are about to forget. Knowing what to teach next is only half the problem. The other half is knowing what you already learned and are at risk of losing, and bringing it back before that happens rather than after.

Put those three together and a system stops recommending "popular next lessons for people like you," which is really just content-based filtering wearing an AI label, and starts picking the one lesson that moves your specific, current gap forward, while also resurfacing what is about to fade. That combination, not the presence of a chatbot, is what personalization actually means.

Where each input comes from, step by step

Step 1: map where you actually start

Before a system can personalize anything, it needs an honest starting point. A generic onboarding form that asks "how would you rate your skills, 1 to 5" produces exactly the unreliable self-rating problem above. A better starting point is a structured assessment, what we call a Learning Scan, that infers your actual starting level from how you respond to real scenarios rather than how you describe yourself.

Step 2: pick today's one lesson

With a starting point and a stated role and goals, a recommender narrows a large content library down to candidates that are relevant to you specifically. A naive version stops here and just returns "the most relevant lesson," which tends to produce narrow, repetitive suggestions: the same three topics in slightly different wrapping, because they scored highest on relevance and nothing pushed back on that. A better version balances relevance against diversity, so the path does not get stuck in a rut, and against your demonstrated mastery, so you are not shown something you have already proven you know or something too far ahead of where you actually are.

We wrote a fuller version of how this path gets built without a human curriculum designer manually assembling it in designing a personalized learning path without an L&D team.

Step 3: estimate what you actually know

This is where Bayesian Knowledge Tracing does its work. BKT is not new or exotic; it comes from intelligent tutoring system research, introduced by Corbett and Anderson in 1994, and has been used in adaptive learning systems for three decades. The core idea: treat mastery of a specific skill as a hidden probability that gets updated with each new piece of evidence (a correct or incorrect response), rather than read off a single quiz score.

The practical effect is that mastery estimates move gradually and account for noise. One lucky guess does not flip your status to "mastered," and one careless mistake on a skill you genuinely know does not flip it back to "not mastered." That is a meaningfully different, and more honest, signal than "percent complete."

Step 4: bring back what is about to fade

Even accurate, demonstrated knowledge decays if it is never revisited. This is the forgetting curve, published by Hermann Ebbinghaus in 1885 and replicated in various forms ever since: unreinforced material fades on a roughly predictable schedule, faster at first, then more slowly.

Spaced repetition systems exploit that predictability on purpose. Instead of reviewing everything on a fixed weekly schedule (wasteful, since well-known material does not need weekly review) or never reviewing at all (the default for most course libraries, since a course typically ends and nothing brings you back), a scheduling algorithm estimates when each specific piece of knowledge is about to cross back below a "likely forgotten" threshold and resurfaces it right before that happens. Omie uses FSRS, a modern open scheduling algorithm built on this research, to decide when old material reappears inside the daily 10 minutes rather than as a separate review chore.

What this actually looks like day to day

For a learner: you open your one lesson for today. It is chosen because it is relevant to your role and goals, sits just past what you have already demonstrated, and the library was checked for diversity so you are not stuck in a rut. Some days, part of that ten minutes is a quick resurfaced check on something from three weeks ago, timed because the system estimates you are about to lose it, not because it is Monday.

For a manager or an HR lead, the same machinery answers a different, longer-standing question. Instead of a completion percentage, a personalization system built this way can report which skills are actually moving, on which people, because the mastery estimate underneath it updates continuously rather than only at the end of a course. That is the deeper argument for why personalization matters at the team level, not just the individual one: it is the same infrastructure that makes proof of skill movement possible instead of a certificate of attendance.

What AI personalization is not

It is not a chatbot bolted onto a course library. A conversational interface is a legitimate front door, useful for asking questions or getting coaching in the moment, but it is not where the personalization work happens. If you removed the chat entirely, an honestly personalized system would still pick a different next lesson for every person, because that decision is made by the recommender and the mastery model, not the chat window.

It is also not general workplace surveillance. The inputs above are learning-specific: what you have done, how you performed on checks tied to that content, and what you told the system about your role and goals. A vendor that cannot say plainly which of the three inputs above it actually uses, and instead answers every question with "our proprietary AI," is worth a more skeptical look.

FAQ

How does AI personalize learning? By combining three inputs per person: role and goals, demonstrated mastery (not a self-rating), and a forgetting model of what is about to fade. A recommender combines them to pick one specific lesson, instead of a fixed playlist for everyone.

What is Bayesian Knowledge Tracing? A method from intelligent tutoring research (Corbett and Anderson, 1994) that estimates the probability you have mastered a skill from your pattern of responses over time, rather than a single quiz score, updating gradually as new evidence comes in.

What does spaced repetition have to do with personalization? It decides when old material comes back, not just what is new. Built on Ebbinghaus's 1885 forgetting curve research: unreinforced material fades on a predictable schedule, and a good system resurfaces it right before that happens.

Is AI personalization just a chatbot? No. A chat interface, where one exists, is the front door. The actual personalization, which lesson you see today and when older material returns, happens in a recommender system behind it, chat window or not.

Does this mean the system tracks everything I do? It tracks learning-specific signals: lessons completed, performance on related checks, and stated role and goals. That is narrower than general surveillance, and a system built honestly can tell you exactly what it uses and why.

Curious what AI-personalized learning looks like for your own role and goals? Run a free Learning Scan →