I spent a good amount of time building what I thought was an ideal AI tutors for my courses. I made it carefully Socratic. I asked it to avoid direct answers, to respond with questions, to nudge students toward their own reasoning. Technically, it functioned just as I designed it. When I used it like a student who is tired and under time pressure, the charm faded quickly. I wanted a clear explanation, and it kept giving me more questions. After another round of tuning, I tried to make it friendlier and more supportive. Students then told me that the tutor felt noisy and overwhelming. At that point I understood that I was training them to handle my special tutor, not the kind of AI they actually meet outside the class. That is a strange educational goal.
Hobbled AI tutors feel safe and ethical, but they miseducate. They train students to work with artificial constraints that disappear the moment they open a normal AI system in a browser. We act as if a restricted tool is a good stepping stone toward a more powerful one. In practice, students build habits that do not transfer. They learn that AI always refuses direct answers, always behaves in a certain tone, always follows classroom rules. Then they encounter a general model that does none of those things, and much of their practice becomes irrelevant.
This is not a new pattern. Education has long relied on simplified versions of reality. We create word problems that clean up numbers, experiments that always work if you follow the manual, and case studies that fit on two pages. Those devices lower risk and cognitive load. They provide a controlled environment where mistakes are safe and visible. The logic is understandable. A student who is still learning should not make an error that costs a patient, a client, or a company. For that reason, some version of a sandbox is necessary.
The trouble appears when we forget that the sandbox is not the field of practice itself. The rules inside a controlled environment do not match the rules in workplaces, in graduate study, or in everyday online life. In many disciplines, educators now try to bring more authentic tasks into courses, so that students face messy data, conflicting evidence, and imperfect instructions. With AI, that tension between safety and authenticity becomes sharper, because the distance between the restricted version and the real tool is very large.
Once we start to protect students by modifying AI itself, we create a peculiar hybrid. We instruct the model never to give full solutions, or to delay any concrete suggestion until several rounds of questions. We narrow its sources and formats. We insist on a very specific teacherly tone that no professional tool will ever reproduce. The model still has the power to generate complex text, but it is prevented from using that power in the ways that matter most outside class. Students feel both burdened and confused. The system is strong enough to dominate the interaction, yet weak enough to be unhelpful when they need efficiency.
I do not think the answer is to abandon scaffolding. It is to move scaffolding out of the model and into our teaching. Instead of hard technical restraints, we can offer social and cognitive guidance. We can talk about appropriate and inappropriate uses of AI for a given assignment. We can model how to break a task into steps and how to design prompts for each step. We can teach students to read AI output with the same suspicion they bring to an unfamiliar website, to check claims against other sources, and to notice when the model clearly fabricates information.
In my own courses, this leads to a split strategy. For routine classroom work, a custom AI assistant still makes sense. It can generate weekly reading lists aligned with the syllabus, create small formative quizzes, and support simple administrative tasks. Those are narrow functions where tight constraints are actually helpful, because students do not need to reuse those bots later in life. They will not need my quiz generator at work.
For substantial projects, I will now invite students to use the same broad AI tools that everyone else uses. I want them to confront vague or partial answers, and to learn how to ask for clarification. I want them to see different versions of an argument and to practice choosing which one is worth pursuing. That means teaching very specific skills. For example, how to ask the model to reveal its uncertainty, how to request alternative lines of reasoning, how to move from a generic first draft to a more precise second version, and how to document their own use of AI in an honest way.
This approach accepts that mistakes will happen. Some students will trust the model too much. Some will misread an answer. But those risks already exist when they use AI on their own phones and laptops, far away from the course platform. In that context, a hobbled classroom tutor does not protect them. It leaves them underprepared. They know how to navigate a special kind of AI that appears only inside one course, and they lack practice with the systems they actually depend on.
An AI tutor that is permanently handicapped may look safer to us as instructors, but it does not prepare students for real learning with AI. It produces clever conversations inside a narrow frame and trains habits that fail outside that frame. I would rather expose students to the real tools and walk with them through the confusion, than give them a polished imitation that vanishes as soon as the course ends.
