What AI Still Can’t Teach

Three gaps AI hasn’t closed yet + why they matter for learning design

Image by ChatGPT

I’ve been watching something quietly frustrating play out in digital learning.

The tools are extraordinary. AI tutors that adapt in real time. Platforms that can spot a struggling learner before their teacher does. Personalized content paths that actually work.

But here’s what else is happening.

A 2025 randomized experiment at a Budapest university tried to do something simple. They started to split students into two groups. One was using AI freely and the other working without it. The study wanted to measure the difference in learning outcomes.

What happened? They never got to finish the study.

Why? Students in the no-AI group rebelled so forcefully that the experiment made national news and escalated all the way to Hungary’s State Secretary for Higher Education. Not because they argued AI made them learn better. But because working without it had become simply unimaginable. The researchers ended up merging the groups before the experiment concluded, but noted that the students’ reaction was itself the most revealing data point. Generative AI had already become so indispensable that being asked to set it aside felt like a violation.

That’s not a technology story. That’s a dependency story.

And yet. AI tools and tutors alone are not the answer. What I’ve seen and heard across the products I’ve been touching is that learners keep dropping out of AI-led learning conversations. They start. They disengage. They don’t come back.

After 25 years designing learning experiences, I think I know why. And it has nothing to do with the technology.

AI is brilliant at teaching when someone already wants to learn.

It reduces the anxiety of being judged. It lets you try and fail privately, at 2am, without embarrassment. For learners who are motivated, organized, and just need access and practice, AI is a genuine breakthrough

image by ChatGPT

The motivational gap

Motivation isn’t a feature you can build into a model. This moment feels remarkably similar to when Massive Open Online Courses (MOOCs), circa 2010, hit the internet. They were all the rage. Millions signed up. The challenge was that fewer than 5% of people actually completed any of these courses. Similar to MOOCs, AI alone remains compelling only for the already highly motivated learner. The first-generation college student using AI tutoring to prepare for the MCAT. The mid-career professional trying to learn data skills to stay relevant at work, squeezing in modules between meetings and school pickups. The person with a clogged sink ay home who desperately needs to fix it themselves.

By far, the vast majority of people don’t live with that level of urgency. They have competing priorities and demands on their time getting in the way of a pure, highly motivated desire to learn.

And most of the people organizations and educational institutions most need to reach, the ones who are behind, overwhelmed, or quietly checked out, they’re the same ones who didn’t thrive in MOOCs. And they aren’t the ones thriving in AI learning environments. Not yet.

The thinking gap

There’s also something harder to name. A few researchers are starting to call it the “cognitive offloading” problem: when AI smooths the path too well, it may be trading speed for depth. Less struggle. Less thinking. In one study, students using AI for academic writing showed significantly lower cognitive engagement than those who worked without it. One paper’s phrase stuck with me: “ChatGPT produces more lazy thinkers.”

That’s not a reason to abandon AI in learning. It’s a design challenge.

The belonging gap

Then there’s the gap I keep coming back to: belonging.

Learning the hard stuff such as changing a belief, building a skill that requires real vulnerability, staying in the discomfort long enough to actually grow. This type of learning is relational. It happens between people. I know from years of work with Education Design Lab that belonging is central to whether learners succeed.

Research is now starting to confirm what practitioners have long known: replacing human interaction with AI conversation can quietly erode the sense of relational belonging that makes hard learning sustainable.

At Supernova, an AI English language tutoring platform in India, CEO Maharishi R B recently spoke about the fact that their platform is heavy with AI tutors, but started adding a human mentor for periodic check-ins. This person is not teaching anything. Just checking in to see how things are going. And when Supernova they did this, retention tripled.

An even more interesting experiment: using AI not to replace the human coach, but to alert them. AI monitors engagement and flags when a learner is drifting, and then a real human shows up to help.

This is the type of experience that feels powerful to me. Blending AI with real human educators. Increasing scale with tech, but not removing human teachers completely.

The question to shape learning design

The question I’m sitting with isn’t whether AI belongs in learning design. It absolutely does. And honestly, AI’s entry into learning feels unstoppable regardless. The real question is: how do we learn equally quickly about what the human still has to hold?

Motivation. Judgment. And the sense of belonging that makes someone willing to keep trying.

Because those are what lead to real learning.

Those aren’t gaps to close. They’re central to the design brief.

What are you seeing in your own work — where is AI in learning landing, and where is it falling flat?

This article was written with research assistance and editing provided by Claude.