14 August 2026 · Alex Abramson

How AI Is Changing Education in 2026

The first thing AI did to education was make text infinitely cheap. Any student can now get a paragraph explaining anything, instantly, at any reading level, at three in the morning. That is a genuine change and it is easy to underrate because it happened so fast.

Three years in, it is also clearly not the whole thing.

What the text era actually solved

The honest case for chat-based AI tutoring is strong. It is available at the moment of confusion, which is when help is worth the most and when a human teacher is least likely to be free. It never gets impatient on the fourth rephrasing. It will explain the same idea at five different levels until one lands. For a student who would otherwise have given up on a problem set at midnight, that is not a small thing.

The limits are equally clear. A language model that has produced a fluent paragraph has produced a fluent paragraph; whether the paragraph is correct is a separate question, and fluency is not evidence. There is also a well-documented gap between feeling like you understood an explanation and being able to reproduce it, and a smooth confident paragraph is very good at producing the first feeling without the second.

And then there is the category of thing that text simply is not the right medium for.

Some explanations do not want to be sentences

Try describing, in words alone, how a Fourier transform decomposes a signal. Or how two DNA strands separate and each templates a new partner. Or why an aeroplane wing generates lift, without hand-waving.

You can do it. Textbooks have done it for a century. But the sentences are doing translation work: the concept is spatial, or temporal, or both, and prose has to serialise it into a line of words that the reader then has to reassemble into a picture. The reassembly is the hard part, and it is where people get lost.

None of that is a failure of AI writing. It is a property of the concept. Some ideas are shaped like pictures, and every time you render one as text somebody has to convert it back.

Where the good tools already are

The existing generation deserves more precision than it usually gets, because "AI will replace it" tends to be wrong in an uninteresting way.

Khan Academy's strength was never the video; it was the structured progression and the practice loop — knowing what you have mastered and what comes next. Duolingo's is spaced repetition and daily habit, which is a psychology problem, not a content problem. 3Blue1Brown, which is closer to what we do, demonstrated something different again: that a genuinely well-designed animation can make a piece of mathematics feel obvious in eight minutes that felt impossible for a term.

None of those is a text problem, and none of them is solved by generating more text faster. What they share is design: someone decided what to show and in what order.

The 2026 shift: generated visuals, not generated words

The change worth paying attention to is that the design step is becoming automatable.

A 3Blue1Brown video takes weeks. Hence a few hundred of them rather than a few million, covering the topics one person found interesting rather than the topic you are stuck on tonight. The bottleneck was never the idea that animation helps. It was that animation is expensive to make, so it only exists for questions popular enough to justify the cost.

If a model can write the program that draws the figure, that constraint changes shape. The explanation for a question nobody has ever asked can be built on demand, for that question, because nothing is being retrieved from a library — it is being computed.

This is a different category from a chatbot, and not simply a faster way to produce paragraphs. It is explanation in the medium the concept actually wants.

What to be sceptical about

Some caution is warranted, including about our own category.

A generated animation can be confidently wrong in exactly the way a generated paragraph can, and a wrong diagram is arguably the more persuasive of the two, because it looks like the output of a calculation. Visuals carry authority they have not necessarily earned.

There is also a real risk of a fluency trap in the other direction. Watching a beautiful animation of a concept is enjoyable, and enjoyment is not comprehension. The research on this is uncomfortable. The study techniques that feel most productive, chiefly rereading and highlighting and watching, are consistently outperformed by the ones that feel like work: retrieval practice and spaced repetition. An animated explanation is a very good way to understand something the first time, and no substitute at all for then trying to explain it back without looking.

The useful framing, we think, is that AI is not replacing the teacher or the textbook. It is removing one specific and long-standing constraint: a custom explanation, drawn for your exact question, used to cost a person several weeks.

Jentoo AI is our attempt at the visual half of that. Ask a question — any subject, or nothing to do with school at all — and it generates an animated explanation with narration, built for that question. It is free to try while we are in beta, and like everything in this category, it is worth checking rather than trusting.

Written by Alex Abramson. Published 14 August 2026.