The Myth Persists

In 2023, Sal Khan gave a TED Talk that was watched by millions. He said we were at "the cusp of using AI for probably the biggest positive transformation that education has ever seen." He wrote a book. He appeared on 60 Minutes. The moment had all the machinery of the OpenAI wave behind it — the timing, the credibility, the production. It traveled far.

I believed him. Not blindly — I'd seen the cycles before — but this time felt different. The technology was genuinely better. The timing felt right. I wanted it to be true.

Recently he gave an interview reflecting on Khanmigo, the AI tutoring assistant Khan Academy built off the back of that moment. He was honest. "For a lot of students, it was a non-event," he said. "They just didn't use it much." His chief learning officer put it plainly: "So far I am not seeing the revolution in education."

That interview didn't travel the same way.

Not because anyone suppressed it. Not because Sal was being evasive — he wasn't. But because the media ecosystem doesn't amplify correction the way it amplifies a promise. A TED Talk about the future of learning in the age of AI is a story. A thoughtful reflection on why the chatbot didn't move the needle the way you hoped is a much quieter one. So most people who saw the first story will never encounter the second. The myth stays intact not through deception, but through asymmetry.

There's a pattern that shows up every few decades. A new technology arrives and someone credible announces it will revolutionize education. Radio. Television. Computers in classrooms. MOOCs. Tablets. And now AI. The Veritasium video about this is worth watching — the same cycle, documented with different faces. The promise is always the same: finally, we can reach everyone, at their pace, with the perfect explanation.

And then it doesn't happen. Or it makes things worse in ways nobody predicted. And then the next technology arrives and we do it again.

Sal Khan is not a fraud. He genuinely believed it. He tried. He put serious resources behind it, and he's been honest about what he found. The problem isn't what he said — it's what gets heard. The original promise is louder than the correction by design. That's how the cycle survives.

What personalized tutoring misses

The promise of AI tutoring is seductive: a patient, always-available tutor that adapts to your level, never gets frustrated, explains it a different way when you don't get it. Better than nothing, for sure. Maybe useful in specific situations. But it doesn't move the needle the way the pitch suggests.

The chatbot interface has real limits that aren't talked about enough. It can only see what you write. It can't tell that you've been staring at the same problem for fifteen minutes and your frustration has flipped from productive to stuck. It can't notice that you're using confident language but making the same mistake repeatedly. It can't decide to pivot to a simpler example, or change the topic for a few minutes, or just acknowledge that this one is genuinely hard.

A good teacher does all of those things without being asked. Not because they have access to a database of pedagogical techniques — because they're reading a human in real time and adjusting continuously. That's not something you solve by giving the AI more context. It's a different kind of perception.

But there's something deeper than perception. A student lets a teacher push them — really push them, past where it's comfortable — because they trust that the teacher is on their side. That trust is built through time, through being known, through the teacher noticing something specific about how this particular person thinks. It's relational. A chatbot can personalize content, but it cannot be trusted the way a person can. And without that trust, the whole dynamic changes. Students don't let themselves be challenged by something that doesn't know them.

This is what personalized learning at scale loses. The 1:1 relationship between a teacher and a student is not just an efficient delivery mechanism — it's the thing itself. You cannot scale that without losing what made it work.

The cheating question

One of the things that gets raised about AI in education is cheating. Students use it to write essays, solve problems, complete assignments. Teachers scramble. Schools ban phones. Institutions panic.

I'm not sure cheating is the right frame.

If a skill becomes something an AI can do reliably on demand, the question worth asking is whether that skill was worth the exercise in the first place. Not all of them have the same answer. Remembering historical dates when you can search them — different from understanding why those events happened. Knowing the multiplication table by heart when you have a calculator — different from understanding what multiplication actually means.

The brain needs reps. I'm not saying don't practice. But the rep should be in service of something. Learn how multiplication works, not just the table. Once you do enough problems, the patterns stick anyway — not because you drilled them, but because you understood enough to recognize them. The principle matters more than the particular algorithm.

So when students use AI to complete an assignment, the right question is: was that assignment developing something real, or was it measuring compliance? Those are not the same thing. A lot of what gets called cheating is students finding efficient routes around work that wasn't going to teach them anything.

That's not a defense of disengagement. It's a question about what we're actually measuring.

What actually works

I keep coming back to one variable: whether the person wants to be there.

When I taught math, the students I remember most were not the ones who had the most natural ability. They were the ones who were curious about it. They stayed after questions were answered. They pushed on the edge of what I'd explained. They came back with something they'd tried on their own. That self-selection changed the entire dynamic. I could push harder, go further, have actual conversations rather than just managing resistance.

Part of it is motivation. If a learner doesn't care, doesn't want to be there, doesn't see the point — a better interface won't fix that. Motivation is upstream of everything. A determined person will learn from a bad teacher, a terrible textbook, a flickering screen. An indifferent person won't learn from the best system you can design.

But even the motivated students — the ones who stayed after class — were not learning from me the way they would have learned from a chatbot. They were learning from someone who knew them. Who remembered what had frustrated them last week. Who adjusted their tone when they sensed confidence was low. Who said "I think you can do this" and had enough history with them that it landed. That's what I was providing, and it had very little to do with my pedagogical technique and everything to do with being a person they trusted.

This is the thing AI tutoring cannot install — not because the technology isn't good enough, but because trust isn't a feature. It's the outcome of a relationship. And a relationship requires time, specificity, and another person on the other end who actually cares about this particular student.

Where AI fits

I don't think AI in education is a dead end. I think it's being evaluated against assumptions that no longer hold.

The assumption is that the bottleneck is access to explanation. If only every student could have a patient tutor available at all times, the problem would be solved. But that's not the bottleneck. The bottleneck is motivation, context, and the relational trust that only forms between two specific people over time.

Plain ChatGPT probably has more impact than Khanmigo, and not because it's technically superior. It's because people who use it are choosing to. Nobody assigned it to them. They came with a question they actually had. That changes everything about how they receive the answer. And when someone chooses to use a tool repeatedly, something like familiarity develops — not trust in the relational sense, but a working fluency that's better than nothing.

Wikipedia was rejected by academics and institutions when it arrived. Now it's the default starting point for most people who want to understand something. It didn't replace books or experts — it became part of the ecosystem in a way nobody predicted. I think AI tools will do the same thing. Not as replacement tutors deployed inside school systems, but as things curious people use because they're useful.

The people who care will find the way around any limitation. The people who don't — no tool designed for them will save them, because the problem was never the tool.

Sal Khan tried. He's been honest about what happened. The correction exists. It just didn't travel.

What I understand now that I didn't in 2023 is what Khanmigo was actually trying to do. It was trying to bring the benefits of 1:1 personalized learning to everyone at once. That's a worthy goal. But the benefits of 1:1 learning are inseparable from the 1:1 part — from the fact that one specific person is paying attention to one specific student, building something over time. The moment you scale that, you've removed the ingredient that made it work.

The TED Talk lives on YouTube with millions of views. The nuanced follow-up lives in a Chalkbeat article. The myth keeps running not because anyone is lying, but because that's how amplification works. Which means the next cycle is already starting somewhere, built on top of a promise that was never fully corrected.

The technology will keep improving. The next pitch will be more convincing. But the relational substrate of real learning — the trust, the specificity, the history between two people — that doesn't get better with a new model. It requires a different kind of investment entirely.