Rethinking ChatGPT for Learning

As of 2026, nine out of ten Norwegian higher-education students have used AI in their studies.1 And worldwide, ChatGPT dominates that use.2

Early evidence of its effect on learning is concerning. In one study, students using out-of-the-box ChatGPT scored 17% worse on exams than students who used no AI at all.3

So ChatGPT, at its worst, isn’t merely unhelpful for students; it’s actively hurting learning.

My goal for this article is to name the specific problems, then show the concrete product changes that would flip ChatGPT from detrimental to beneficial for students.

The problem with ChatGPT

Assistant, not tutor. ChatGPT is designed to follow instructions, complete the task, and appease the user. A good tutor often has to do the opposite: withhold the answer, push back, and make the student work through the problem. ChatGPT is also fundamentally reactive. It answers the latest prompt, then waits. A good tutor does not merely respond; they steer the whole interaction. They already have a sense of what the student is working toward and how each step fits into a larger learning path, so they can guide the conversation rather than passively follow along. For most work, the output is the goal. In education, how the student reaches the output is often the whole point.

Consuming, not learning. That ask-and-answer pattern makes it easy to confuse receiving an explanation with learning. There is no built-in requirement for the student to retrieve the information, apply it, or demonstrate that they understood it. There is also no feedback loop through which the AI can verify that its teaching worked. The result is an interaction built around consuming information, even though durable learning requires the student to actively engage with the material.

Continuity and grounding. In formal education, learning is usually structured as an overarching goal broken into sub-parts, mastered across coordinated sessions that add up to the end goal. ChatGPT may remember fragments about a user, but that is not the same as maintaining an explicit, continuously updated model of what the student knows, misunderstands, and is ready to learn next. Without that, it cannot reliably orchestrate a path across sessions. It also lacks the local context that shapes real education: the student’s course, textbook, assignments, teacher, and previous work. That leaves the student to keep pasting in what matters, instead of the AI already knowing it. Without that grounding, it can personalize answers only superficially. The student has to continuously re-explain their world. The AI never really shares it.

Interface and medium. The blank chatbot screen places all the initiative on the user. They have to know what to ask, how to ask it, and what they should do next, precisely what they often need help figuring out. Education requires more scaffolding than an empty text box provides. The interaction also defaults to written conversation when another medium might teach better: a diagram, simulation, quiz, game, podcast, or video.

Why ChatGPT won’t fix this

These issues are unlikely to be solved by ChatGPT. The primary reason is that an educational tool needs specialized skills and design, and those require trade-offs that are just not worth it for a general-purpose product. One of the clearest clashes is between specializing for education and for work. They are completely at odds with each other. It's clear from the financial incentives perspective that they will always choose the latter.

They can make adjustments around the edges, such as introducing a dedicated study mode whose main change is that it doesn’t give answers instantly. That’s obviously better than the default, but it’s not enough to solve the underlying problems.

Shape of the solution

The north star is to help students attain durable knowledge that they can apply independently.

That requires a solution that sees its role as not relieving the student of their thinking but guiding it, that incorporates pedagogical best practices, that gets the student learning by actively producing, not passively consuming, information.

It has to be clear-eyed about student psychology: students often lean toward what is easy and convenient. It has a responsibility to actively steer them toward their long-term educational success.

It needs to carefully balance what AI should help with and what it should not. Take the gym as an analogy. You could superficially observe it and conclude that the way to help people is to build a forklift that carries the weights for them, mistaking the activity for the underlying goal.

2x solution

My solution within the chatbot paradigm has two parts: architecture and capabilities. They solve different problems, but together they make a coherent AI tutor.

Architecture. The goal is for the AI tutor to interact with and understand the student as a human tutor would.

• Proactive. Move away from the reactive default of chatbots to a proactive style that drives the conversation the way a human tutor would, monitoring the student’s activity and intervening based on learning signals. The simplest way is to introduce session state. Always include a snippet in the context with the up-to-date objective and sub-objectives. That gives it directionality to break subjects into steps and operate above turn-by-turn reacting.
• Always verify knowledge. To solve the issue of students mostly passively consuming information, students should always have to prove that they have learned the material before moving on. You do this by implementing as many feedback loops as possible, forcing the students to reproduce the material in different ways. The richer the feedback loop, the better, so the most basic feedback would be taking a multiple-choice question, for example, and a richer version would be demonstrating mastery in more specific ways, like for coding, writing an actual program.
• Above session layer. Sessions shouldn’t be siloed one-offs. You need a layer above them that tracks every session, manages a durable knowledge state, and runs multi-session plans the way real tutoring works. It should also be able to communicate directly with the student outside student-initiated conversations.
• Pedagogical best practices. How to teach effectively is well known. Put those fundamentals in the tutor’s instructions: desirable difficulty, mastery learning, zone of proximal development, productive struggle, etc. That way the AI actually teach well.
• Knowledge state. A human tutor builds a mental model of the student: their learning style and their pre-existing knowledge. That data is critical for tailoring and driving the tutoring. To get it, the AI needs a knowledge state system above the sessions that collects this over time and synthesizes it.

Capabilities. The goal is for the AI tutor to act on that understanding as a human tutor would.

• Teaching. To teach effectively, we need to move beyond plain text answers. First, mimic the abilities of the human tutor: give the AI ways to explain visually, the equivalent of giving a tutor pen and paper. Second, move past the limitations of a human tutor and create tailored learning material like podcasts, videos, games, and interactive simulations.
• Testing. Verifying that the teaching has successfully been understood is critical. To be able to do that, we need specialized tools whose sole purpose is to test the student. This should vary from very simple one-off questions to more elaborate exam-like tests. It should incorporate all methods like written, visual, and oral to match the diverse ways students get tested in a school environment.
• Context. AI is always bottlenecked by the quality of its context. So we need specialized context-gathering tools: the exact textbook or literature the student is using, and ways for students to reference their own material. Think of it as giving the AI tutor eyes and ears. With a human tutor, a student can simply point at something and the tutor picks it up from the shared environment. The same ability, for AI.

Properly integrated, I believe these would be highly valuable for students. But they’re still trapped in the chatbot paradigm, which caps how far you can go.

10x solution

To reach the next level, a step-change improvement, you need to break through the chatbot interface paradigm and build something from the ground up, exclusively for education.

I wrote a full piece just on this. Instead of reiterating it here, I recommend reading that one.

EduOSHow I would rebuild education around AI

My work

I currently work on an exploratory project called Sixtus where I try to implement some of the things outlined in the 2x solution block.

I open-source all my code on GitHub, so check that out if you're interested in the technical implementation.

Sources

1. Studiebarometeret 2025 — Norwegian Directorate for Higher Education and Skills. Survey finding: 91% of Norwegian higher-education students had used generative AI in their studies.
2. Global AI Student Survey 2024 — Digital Education Council. International survey finding: ChatGPT is the most widely used AI tool among students.
3. Generative AI without guardrails can harm learning: Evidence from high school mathematics — Bastani et al., PNAS. Turkish high school math study: students using base ChatGPT scored worse on exams than students with no AI.