A student opens an AI tutor, types a question, and gets a clear, fluent answer in seconds. They nod along. They feel like they understand. They close the laptop. The next day, on a quiz, they can’t reproduce a single step of the reasoning. What happened?
This isn’t a story about bad students. It’s a story about how a generation of learners is being shaped by tools that are extraordinarily good at producing the feeling of understanding, without necessarily producing understanding itself.
The fluency trap
Large language models are, in a literal sense, fluency machines. They are trained to produce text that flows naturally, sounds confident, and reads as if it knows what it’s talking about. For learners, this fluency is intoxicating. It creates what cognitive scientists call the illusion of explanatory depth, the sense that we understand something deeply because the explanation we just read sounded coherent.
In conversations with EdTech founders across 21 countries, this concern surfaced again and again. As one founder put it:
When the scaffold becomes a crutch
Good scaffolding in education is designed to fade. A teacher gives a hint, then a smaller hint, then steps back entirely. The point is to support the learner just past the edge of what they can do alone, and then to remove that support as soon as they can stand on their own.
AI tutors, by contrast, are usually designed to be infinitely patient and infinitely available. They never get tired of giving hints. They never push back. They never refuse to help. And because of that, the scaffolding never fades.
One founder asked the central question directly: “When does the scaffold become a crutch? When does having access to these scaffolds actually limit student learning?”
The convincingly wrong problem
There’s a second layer to this. AI doesn’t just produce convincing-sounding right answers. It produces convincing-sounding wrong answers with the same confidence and the same fluency. For a novice learner, someone who, by definition, doesn’t yet know enough to evaluate the answer, this is a serious problem.
What real learning looks like
Real learning is uncomfortable. It requires what researchers call “productive struggle”, the messy, frustrating work of grappling with an idea you don’t yet understand until something clicks. It requires generating your own explanations, not consuming someone else’s. It requires being wrong, noticing you were wrong, and figuring out why.
None of that is fast. None of it feels good in the moment. And none of it shows up on a dashboard as “minutes engaged with the platform.”
This is the design challenge facing the next generation of EdTech: how do we build AI that supports the discomfort of real learning rather than dissolving it?
Designing for depth, not just fluency
Some founders are already experimenting with answers. Socratic tutors that ask questions instead of giving them. Hints that are deliberately delayed. Scaffolds that fade automatically over time. Prompts that ask students to explain their reasoning before the AI responds. Verification mechanisms that surface uncertainty rather than hiding it.
None of these are perfect. But they share a common premise: the goal of an educational AI is not to maximize the appearance of learning. It is to maximize the conditions under which learning actually happens.
That is a harder product to build. It is a harder product to sell. It will not always look as impressive in a demo. But it is the only product worth building.