AI Can Teach Anyone. The Hard Part Is Teaching You.

Almost everyone got a brilliant private tutor in the last two years. It sits in a browser tab, never sleeps, and will explain the French subjunctive or the second law of thermodynamics or a nasty logic puzzle at two in the morning without ever getting bored of you. For the price of a subscription, or for nothing at all, anyone can now ask an infinitely patient expert to rephrase an idea until it finally lands.

So here is the uncomfortable question: if everyone now has a genius on call, why isn't everyone learning faster?

This is a genuinely big deal, and it is easy to underrate because it arrived so fast. For most of history the bottleneck in learning was access: to a teacher, to an explanation, to someone who could answer the exact question stuck in your head right now. That bottleneck is mostly gone.

A large language model is astonishing at explaining things and strangely limited at teaching a particular person, because explaining and teaching are not the same job.

Explaining is about the material. Teaching is about the student. And the second one only gets interesting when the software actually knows who you are.

Where AI tutors genuinely help

AI is extraordinary at explanation on demand. You can paste in a paragraph you do not understand and get it unpacked five different ways until one of them clicks. You can ask the same question three times with no social cost. For a nervous learner that absence of judgment is not a minor feature: plenty of people never ask the human teacher the "dumb" question, and the machine dissolves that hesitation entirely.

It is also a tireless generator. Give it a topic and it produces examples, analogies, practice problems, and counterexamples faster than you can read them. It shifts registers on request, from a textbook definition to a plain-English one to a version aimed at a ten-year-old, and it holds a back-and-forth that feels far closer to a conversation than a search box ever did.

If your problem is "I do not understand this thing and there is no one around to ask," AI has quietly solved it. The point of the criticism that follows is not that AI is weak; it is that AI is powerful along one axis and nearly absent along another.

Where it quietly fails

Photo by Milad Fakurian on Unsplash

Open a fresh chat and you are a stranger to it. It has no idea what you already know, what you keep getting wrong, how fast you move, when your exam is, or what you studied yesterday and forgot by this morning. Every session begins at zero.

That single fact produces most of the failure modes.

It cannot tell your strong topics from your weak ones, so it spends the same care on the thing you have mastered as on the thing that keeps tripping you. You can tell it, of course. But a blank chat hands the diagnosis to the one person who cannot do it: the learner, who by definition cannot yet see the shape of their own gaps. If you understood your weaknesses well enough to describe them, you would already be halfway to fixing them.

It also leans toward the satisfying answer over the durable one. A chatbot is tuned to produce a reply you find helpful in the moment. Learning often needs the opposite: being made to retrieve something before you are shown it, being told to stop and review a thing you thought you had finished, being corrected in a way that stings a little. A model with no stake in your long-term progress will happily walk you through the same concept for the hundredth time and never once tell you to move on.

And it can be wrong with total fluency. A confident, well-formatted explanation that is subtly incorrect is more dangerous in learning than an obvious error, because the learner has no independent way to catch it. Add no memory of you, no spaced review, and no ongoing accountability, and you get something that feels like a tutor and behaves like a very articulate book. It will answer anything. It will remember nothing about you.

Where personalized, adaptive learning kicks in

Photo by Growtika on Unsplash

Here is the shift that makes the whole category interesting. Take the same language model, and instead of running it against a blank slate, run it against a persistent picture of the individual learner. Not "here is a question," but "here is this person's last two hundred answers, the topics they own, the ones they fumble, the things they learned three weeks ago and are about to lose, the pace they can sustain, and the deadline they are working toward." Then let the model reason from that.

That single change flips the technology from an answer engine into something that behaves like a coach.

It can notice you have missed four questions in a row that shared a single hidden feature, and name the pattern you could not see yourself. It can decide that today is the exact right day to resurface a concept you are about to forget, not because a calendar says so but because your own recall history says so. It can look at a mistake you just made and hand you the one lesson that speaks to it, instead of a generic list of tips that ignores the mistake you actually made. It can tune the daily workload to the twenty minutes you actually have rather than an idealized syllabus.

None of that comes from a bigger model. It comes from memory, and from the discipline to act on it. Capability and personalization are separate axes. You can have enormous capability with zero personalization, which is exactly what a fresh chat window is. The products that matter are building the second axis, a structured, evolving model of the learner, and pointing the language model at it. This is what the field calls adaptive learning: difficulty, timing, and content that adjust to one person rather than to an average.

Learning science has said this for decades. Spaced repetition works, but only if something tracks when you personally saw a thing and how well you recalled it. Retrieval practice works, but only if something knows what is worth making you retrieve. Targeted practice beats sheer volume, but only if something is tracking, item by item, where your volume should go. Every one of those is a personalization problem dressed in learning-science language, and a generic tool cannot do any of them, because it does not hold your record.

The AI tutors already working this way: Duolingo, Khan Academy, Khanmigo

This is not a thought experiment. Several large, credible products have been building exactly this pattern for years: a rented, general-purpose language model sitting on top of a proprietary, per-user model of the learner.

Duolingo is the cleanest example at scale. Long before the current AI wave, it built a machine-learning system it calls Birdbrain, whose entire job is to predict how likely a specific user is to get a given exercise right and tune difficulty accordingly. That is adaptive learning at the core of the product, not bolted on. When large language models arrived, Duolingo layered them on top through its Max tier, with features like "Explain My Answer," which breaks down why your specific mistake was a mistake, and "Roleplay," a practice conversation with an AI character. The stack is the interesting bit: the language model handles fluent conversation, and the per-learner model underneath handles what you personally need next.

Khan Academy took the same bet from the tutoring direction. Its assistant, Khanmigo, is built on frontier models but deliberately integrated into the student's actual coursework and Khan's mastery system, and it is designed to be Socratic. Instead of handing over the answer, it nudges, asks what you are thinking, and guides you to work it out, which is much closer to what a good human tutor does than a chatbot that just resolves your question and moves on. The design choice is the tell: Khan did not drop a chat box onto the site and call it a tutor. They connected the model to the learner's context on purpose, because that context is where the teaching value lives.

It is worth noticing that even the frontier labs are converging here. General assistants have been adding persistent memory across conversations, precisely because a tool that forgets you every session leaves most of its value on the table. When the companies building the raw models start racing to remember their users, that tells you plainly where the leverage sits.

How to spot the difference

Look across those examples and the common thread is not the model. Everyone is renting broadly similar language models from the same handful of labs. What separates a real learning product from a wrapper around a chatbot is the proprietary, per-user data underneath, and what it does with it. In this category, the moat is memory, not the model.

That reframes the question worth asking, whether you are a learner choosing a tool or a builder making one. Not "does this use the latest AI." Almost everything does now, and it tells you very little. The better question: does this remember me, and does it do something specific with what it remembers?

There is a quick test. When a product tells you what to study next, ask where that recommendation came from. If it traces back to your own record (the questions you missed, the timing of your reviews, the specific things you keep getting wrong), you are looking at a system that actually works with you. If it is a best-practice list that would read identically for every user on the planet, you have a capable model with amnesia, and no amount of model upgrades will fix that, because the missing ingredient was never intelligence.

You can already see the pattern spreading from the big platforms into narrower, higher-stakes niches, the places where knowing the individual learner matters most. In test prep, for example, an app called Grezi applies the same idea to GRE Verbal: its in-app tutor reads your practice history, the words you have seen and the question types you keep missing, and schedules each word's review from your recall rather than a fixed calendar. It is a niche example next to Duolingo's scale, but the architecture is identical: a rented model on top of a per-user record, with the record calling the shots.

The AI that can explain anything already exists, and it is genuinely wonderful for what it is. Keep it open in a tab. Ask it to unpack the hard idea, to rephrase until it clicks, to break down the question that stumped you. Just do not mistake it for something that is teaching you, specifically, on purpose, over time.

Personal is the whole game.

Disclosure: the author is affiliated with Grezi.