AI as a Tutor, Not an Answer Machine

In the previous post, I focused on productive human-AI collaboration. AI can generate possibilities, offer explanations, identify patterns, and challenge assumptions, but students still need to direct the work, evaluate what the tool contributes, and remain responsible for the decisions that follow.

One of the most promising forms of that collaboration is using AI as a tutor. The appeal is easy to understand. Students can ask questions at any time. They can request another explanation without worrying that they are taking up too much of someone’s time. They can ask for examples, practice questions, hints, feedback, or a different way of approaching a concept.

They can tell AI that they are confused. They can ask what a term means. They can ask the same question five different ways. That kind of immediate, responsive support has real potential. But asking AI a question does not automatically make the interaction tutoring. Sometimes it is simply answer delivery.

That matters because an answer machine and a tutor may respond to the same student in very different ways. An answer machine asks, “What response will satisfy this request?” A tutor asks, “What does this student need to do next in order to learn?” Those are not the same goal.

Imagine that a student is working through a networking problem and cannot determine why one computer is unable to access the internet. The student asks AI, “What is wrong?” An answer machine may review the symptoms, identify a likely cause, and provide a list of steps to fix it. The response may be accurate. It may even solve the problem. But what did the student learn?

A tutor might begin by asking what the student has already checked. Can the computer communicate with other devices on the local network? Does it have a valid IP address? Can it reach the default gateway? Can it resolve a domain name? What evidence supports the student’s current diagnosis? Instead of removing the problem, the tutor helps the student work through it. The difference is not that a tutor refuses to help. The difference is the form that help takes.

A good tutor does not simply know the answer. A good tutor makes decisions about when to explain, when to ask a question, when to offer a hint, when to provide an example, and when to let the learner struggle a little longer. That last part can be uncomfortable.

When a student is stuck, giving the answer feels helpful. It reduces frustration and allows the student to move forward. But moving forward in the assignment is not always the same as moving forward in learning. Sometimes the moment of difficulty is where the most important thinking is about to happen.

The student may need to retrieve prior knowledge, test an assumption, compare possibilities, or recognize a gap in their understanding. If AI provides the answer immediately, the difficulty disappears before the student has a chance to work through it. The task is completed. The learning opportunity may not be.

This does not mean struggle is always productive. A student who lacks essential background knowledge may become confused, discouraged, or completely stuck. Repeated failure without useful support does not automatically create learning.

The goal is not maximum difficulty. The goal is appropriate support. A tutor should help students make progress without making the progress for them. That may involve offering one hint at a time. It may involve breaking a complex problem into smaller parts. It may involve asking the student to explain what they already understand. It may involve showing a similar example rather than solving the exact problem. It may involve identifying the step where the student’s reasoning began to go wrong. It may also involve providing a direct explanation when the student genuinely needs one.

The question is not whether the tutor gave help. The question is whether the help preserved the thinking the student needed to practice.

Recent research on generative AI tutoring illustrates this distinction. In a 2025 field experiment involving nearly 1,000 high-school mathematics students, Hamsa Bastani and colleagues compared students who practiced without AI, students with access to a standard GPT-4 interface, and students using a more carefully structured AI tutor.

Students with unrestricted AI access performed better while the tool was available. But when AI was removed during an assessment, they performed worse than students who had practiced without it. The tool helped them reach answers during practice. It did not necessarily help them develop the same ability to reach those answers independently.

The structured tutor produced a different pattern. It was designed to ask students to show their work, identify where they were stuck, and provide guidance without immediately giving away the complete solution. The AI was still useful. But the guardrails changed the kind of help it provided.

Other studies have found promising results from AI tutors designed around established teaching practices. A 2025 randomized study by Greg Kestin and colleagues compared a carefully developed AI tutor with in-class active-learning lessons in an undergraduate physics course. Within the specific lessons studied, students using the AI tutor learned more in less time and reported feeling engaged and motivated.

That result deserves attention. But the important lesson is not simply that an AI chatbot can replace a class. The tutor was deliberately designed around educational principles. It included structured explanations, questions, practice, feedback, and controls intended to keep students engaged with the material.

The learning did not emerge merely because students had access to generative AI. It emerged from how that access was designed.

Research on Tutor CoPilot offers another example. Instead of replacing human tutors, the system provided real-time suggestions to tutors working with students. In a large randomized trial, students whose tutors had access to the system were more likely to master mathematics topics. The analysis also found that supported tutors used more probing questions and were less likely to give away answers.

That is an important model of human-AI collaboration. AI did not become the tutor. AI helped a human tutor make stronger instructional choices. Together, these studies point toward a larger principle. AI tutoring is most promising when it is designed to support learning behaviors, not simply accelerate answer production.

That difference can also be shaped by the student. Most students will not be using a carefully designed experimental tutoring system. They will be using a general-purpose AI tool that is usually willing to provide whatever response they request.

If the student asks for the answer, the tool will often provide it. If the student asks for a complete solution, the tool may provide that too. If the student wants to avoid the difficult part of the task, the system is unlikely to consistently stop them. This means students need to learn how to establish the tutoring relationship themselves.

They might begin with a prompt such as:

“I am learning this topic. Do not give me the answer immediately. Begin by asking what I already understand. Give me one question or hint at a time. Ask me to explain my reasoning before you respond. If I make a mistake, help me identify where my reasoning went wrong rather than simply correcting it.”

That prompt does not guarantee effective tutoring. AI may still misunderstand the student, provide an inaccurate explanation, or reveal too much. But it establishes a different goal for the interaction. The student is not asking AI to complete the problem. The student is asking AI to help them remain engaged with it.

That distinction should continue throughout the conversation. Students can ask AI to question them rather than lecture them. They can ask it to provide examples and then have them generate the next example. They can ask for practice problems without immediately requesting the solutions. They can explain a concept in their own words and ask AI to identify what appears incomplete. They can ask for a hint and then attempt the next step independently. They can request that the difficulty gradually increase as they demonstrate understanding. They can ask AI to return to a concept later to see whether they can retrieve it without assistance.

These interactions position the student as an active learner. That active role is important because explanations can create an illusion of learning. AI can make a difficult concept feel clear. It can produce analogies, organize the information, and remove confusing terminology.

That clarity can be useful. But understanding an explanation while it is in front of us is not the same as being able to explain the concept later. Following a worked solution is not the same as solving a new problem. Recognizing a correct answer is not the same as retrieving it independently.

Students need opportunities to find out whether the explanation became their understanding. After receiving an explanation, they might close or minimize the response and explain the concept in their own words. They might apply the idea to a different situation. They might generate an example that was not included in the AI response. They might solve a similar problem without assistance. They might identify which part they can explain confidently and which part remains unclear.

A useful AI tutor should not only help students understand while the tool is present. It should help them discover what they can do when the tool is absent.

That suggests a simple test. Can I explain this without looking at the response? Can I apply it to a new situation? Can I recognize when it does not apply? Can I solve a similar problem without asking for the same help?

If the answer is no, the tutoring interaction may not be finished. AI tutoring also has another limitation. The system may appear to know the student better than it actually does. Because AI can respond conversationally, refer to earlier statements, and adjust the difficulty of its language, students may feel that it understands their learning needs.

But an AI system does not necessarily possess a reliable model of what the student knows. It may infer misunderstanding from a poorly worded question. It may accept a student’s confident but inaccurate explanation. It may move ahead too quickly. It may repeat the same type of explanation without recognizing that the student needs a different approach. It may provide praise that sounds encouraging but is not connected to real evidence of progress. It may also generate a convincing explanation that is wrong.

A human instructor can make mistakes too. But a human instructor may bring knowledge of the course, the assignment, the student’s previous work, common misconceptions within the discipline, and the broader learning goals. A general-purpose AI system may not have that context unless someone provides it. Even then, it may not use the information reliably.

This is why AI tutoring should not be confused with human understanding. The conversation may feel personal. The response may feel responsive. The tool may remember details within the interaction. But students still need to verify explanations, monitor their own understanding, and seek human support when the problem involves persistent confusion, consequential decisions, or needs the AI cannot adequately recognize.

AI can supplement human support. It should not automatically displace it. There is also a relational side to tutoring that should not be overlooked. A good tutor does more than deliver information. A tutor notices frustration, builds confidence, communicates belief in the learner, and adjusts support based on more than the words in a single response.

A teacher or tutor may recognize that a student’s question is not really about the current problem. The student may lack an earlier concept, fear making a mistake, or need reassurance that difficulty is a normal part of learning.

AI can produce encouraging language. But generated encouragement is not the same as a person recognizing the learner, understanding their history, and choosing to remain present in the struggle with them. That does not make AI tutoring worthless. It helps define its role.

AI may be particularly useful for low-stakes practice, immediate explanations, question generation, guided review, and opportunities to learn outside normal class hours. It may help students prepare better questions before meeting with an instructor. It may help them identify where they are confused. It may provide additional practice when human support is unavailable. It may give students a private space to admit that they do not understand something.

Those are meaningful benefits. But the measure of success should not be how quickly AI helps students finish. It should be whether students become more capable of continuing without it.

This means effective tutoring should gradually reduce support. At first, a student may need a detailed explanation and a worked example. Next, they may need a partial example or several hints. Later, they may need only a question that points them in the right direction. Eventually, they should be able to perform the task independently and explain why their approach works.

If the student always needs the same level of AI support, tutoring may have become dependency. The goal of tutoring is not permanent assistance. The goal is growth. That growth can be made visible through small reflections. What could I do before the AI interaction? What can I do now? What explanation helped? Where did I still need a hint? What mistake did I recognize? What can I now explain without assistance? What should I practice next? These questions shift attention away from whether AI produced a helpful response. They focus on whether the student changed because of the interaction.

That is the standard we should apply to AI tutoring. Did the tool help the student ask a better question? Did it help them retrieve prior knowledge? Did it help them recognize a misconception? Did it offer enough support to make progress without removing the thinking? Did it help the student become more independent? If so, AI may be serving as a tutor.

If it simply provides the answer, completes the steps, and moves the student to the next task, it is functioning as something else. Perhaps a very efficient answer machine. That may be useful in some situations. Not every interaction with AI has to be a learning experience. But when learning is the goal, efficiency cannot be the only measure.

A tutor should not merely make difficult work disappear. A tutor should help students become capable of doing difficult work. One way AI can do that is by providing feedback on something the student has already attempted. Instead of generating the initial work, AI can help students identify weaknesses, question assumptions, and consider possibilities for revision.

But feedback introduces its own questions. Should students accept every suggestion? Can AI critique a piece of work without quietly rewriting it? How can students retain ownership of their decisions while still benefiting from immediate feedback?

That will be the focus of the next post: AI as a critique partner.

Continuing the Conversation

Series 1: AI Is Exposing Existing Problems ✓ Completed
Series 2: What We Do About It ✓ Completed
Series 3: Cultivating Human Thinking ✓ Completed
Series 4: Learning Alongside AI
Current Post (6 of 8): AI as a Tutor, Not an Answer Machine
Next Up: AI as a Critique Partner