Why Real, Widespread AI in Education Products Won’t Happen Anytime Soon

Why Real, Widespread AI in Education Products Won’t Happen Anytime Soon

Not for at least five years.

And even that — only if edtech companies start moving right now.

They won’t.

Thesis 1.

The three most important components in solving any problem with AI are: context, context, and context.

Anyone who uses neural networks a lot knows this: a strong context with a weak prompt gives a better result than a strong prompt with a weak context.

So what is context? Imagine you’re giving a task to a stranger. Context is everything they need to know to complete the task the way you expect — not the way they decide to.

It’s funny, by the way, that people always ask for a prompt and almost never ask, “How did you build the context?”

One reason, I think, is a well-known cognitive bias: “I assume that what I know, everyone else knows too.”

In other words, people can’t even pass context clearly to other people — how are they supposed to pass it to algorithms?

— “Remember that movie where the guy…”

— “???”

— “You know, that movie! Are you dumb? We watched it together!”

That’s roughly how (it seems) 90% of people talk to themselves. That’s how they talk to neural networks.

And that’s also how companies think about context and its role.

So where is the sensible result supposed to come from?

Thesis 2.

Context = data. A lot of data.

And (almost) no one is collecting it.

If you want AI to do something personalized for a student — the most common request I hear — you need to feed it data about that student.

For example: Let’s revoice a TV show into English according to a person’s vocabulary level — they’ll understand almost everything, stay motivated, and churn will drop.

Sure, let’s do that. Technically, it’s easy.

But where will you get that person’s vocabulary? You don’t know it. At best, you can make a list of words they saw in class or homework. But that’s not a vocabulary.

The necessary data here is the words they actually used over a period of time.

And who’s collecting that in education?

Then people say: Let’s analyze students’ mistakes.

Okay. But first, you need a taxonomy of knowledge, map all materials to it, and then collect and store every single mistake the student made — carefully and systematically.

And, again, there are fewer companies doing this on the market than you have fingers on one hand.

Thesis 3.

It’s never too late to start collecting data.

But to do it, you either have to build your own platform or rewrite an existing one — where nothing is being collected at all.

That’s expensive.

And the market right now is stagnating, if not shrinking.

So there’s no money for that. Almost nowhere.

Because we’re in a global investment winter.

Theses 1 + 2 + 3 = poof.

No data (no context) — no real results.

Collecting data is expensive.

And there’s no money.

So, yes, there will be plenty of “AI” — in marketing materials.

Banners will say all the right things.

But in practice, nothing truly cool will appear.

In sales and marketing — yes.

In generating fluff content — sure.

But in the product itself — no.

Except, maybe, for a handful of companies that started thinking about this long ago and have been quietly collecting context for their own purposes.