Standard AI chatbots are too helpful to be good teachers. To protect a child’s ability to think critically, AI must be broken into "modules" that force students to struggle through problems rather than just receiving answers. If a tool makes homework feel effortless, your child likely isn't learning.
Standard AI chatbots prioritize speed and "helpfulness," which often results in giving students the final answer rather than teaching them the process. Researchers propose a "modular" AI tutor that separates the "thinking" from the "answering," ensuring the AI follows proven teaching principles like breaking problems into steps and providing hints rather than solutions.
Convenience is the enemy of deep learning. When your child uses a standard chatbot for a history essay or a math set, the AI’s primary goal is to satisfy the user with a quick, correct response. This creates a "learned helplessness" where the student becomes a prompt-engineer rather than a subject-matter expert.
This finding changes the "AI as a tutor" conversation from if we should use it to how it must be built. If the tools your child uses aren't designed with "desirable difficulty"—the psychological requirement that learning should be a little bit hard—they are essentially just high-tech ways to skip the education you’re paying for.
Researchers are identifying a "pedagogical gap" in modern Large Language Models (LLMs). Current AI models like ChatGPT are built as monolithic "black boxes"—you put a question in, and an answer comes out. There is no internal mechanism that says, "Wait, I shouldn't give the answer yet; I should ask a leading question first."
The current trend in AI development focuses on "frictionless" user experiences. However, education is fundamentally about friction. By attempting to bridge this gap, researchers are trying to move away from "answer machines" toward "agentic tutors" that can actually simulate the behavior of a human teacher who knows when to hold back.
The research team outlines a framework designed to insert intentional "speed bumps" into the AI's logic. They argue that standard, all-in-one models fail at teaching because they lack pedagogical guardrails.
- The creativity trap: Using "all-in-one" AI for homework risks a permanent loss of student creativity and the ability to transfer knowledge to new, slightly different problems.
- Modular architecture: By breaking the AI into specialized "modules," developers can insert specific teaching rules—like "never give the answer on the first try"—at different stages of the conversation.
- Transparency: A modular system makes the machine's "thought process" transparent. This allows parents and educators to see exactly why the AI gave a certain hint or held back certain information, rather than just guessing what's happening inside the software.
There is a fundamental conflict of interest between AI companies and educators. Big Tech companies want "high engagement," and nothing drives engagement like a tool that makes a difficult task (like homework) disappear in seconds. But a tool that makes the work go away also makes the learning go away.
We are currently in a "Wild West" where the AI tools marketed to students are fundamentally designed to do their work for them. This research suggests that until we move toward modular, "agentic" designs, the AI is effectively acting as a high-powered cheat sheet rather than a tutor. We are effectively handing kids a calculator that also writes the word problems, which could result in a generation that knows how to get results but doesn't understand the underlying logic.
This paper is a conceptual architecture proposal, not a study of student performance in a classroom. The authors have designed the how, but they haven't yet proven that kids using this modular system actually score higher on tests than those using standard ChatGPT.
Furthermore, the paper is a preprint and has not yet undergone formal peer review. Its effectiveness depends entirely on the quality of the pedagogical instructions given to each module. If the "teacher" module is programmed with poor teaching strategies, the modularity won't save the student from a subpar learning experience.
- If your child is using ChatGPT for homework... insist they use a "Socratic prompt" to start the session, such as: "Act as a tutor who never gives the answer but only asks guiding questions to help me find it myself."
- If you are evaluating a new educational app... look for "step-by-step" or "hints-only" settings in the menu rather than a wide-open chat interface that allows for copy-pasting.
- If the AI provides a full answer too quickly... treat it as a "spoiler." Ask your child to explain the logic behind the AI's answer back to you in their own words to ensure the knowledge actually transferred from the screen to their brain.
- If you are a teacher or school administrator... prioritize AI tools that advertise "agentic" or "modular" designs, as these are more likely to align with actual classroom goals than general-purpose chat tools.
The goal of educational AI shouldn't be to make homework easier, but to make learning more effective. Until modular tutors that prioritize "desirable difficulty" become the industry standard, parents must act as the pedagogical guardrails that the software currently lacks.
Julius Gabelmann, Felix Jahn, Kevin Baum et al. (2026). Modularizing Educational LLM-Agency for Fostering Responsible Learning Assistance. arXiv (preprint). — arxiv.org