AI Tutors That Don't Do Your Homework: What to Look For

What is an AI tutor that doesn't do your homework?

It is an AI chatbot set up to teach you, not to produce work for you. A research team at Purdue University calls the principle "tutor, not solver": default to hints and scaffolding, and avoid direct answers.

A general chatbot tends to hand you the finished answer. A tutor-style AI changes four behaviours:

• It asks before it answers. Which course is this for, what have you tried, and where are you stuck?
• It gives hints and steps, not the final answer. One step or guiding question at a time, then it waits for your attempt.
• It checks your understanding. You explain the idea back or try a similar problem alone.
• It won't produce work you would submit. No finished essays, solutions or lab reports for graded tasks.

The aim is not a weaker AI. It is keeping the thinking, the part that builds skill, on your side of the screen.

Does the design really matter? What the research says

Yes. Two randomized studies suggest an AI tutor's design changes whether students learn from it, though both have real limits.

The high school math study (Bastani et al., PNAS)

In a 2025 PNAS paper titled "Generative AI without guardrails can harm learning", Hamsa Bastani and colleagues gave nearly 1,000 high school math students in Turkey access to GPT-4 during practice sessions. One group used "GPT Base", which mimicked ChatGPT. Another used "GPT Tutor", prompted to give teacher-designed hints instead of answers and supplied with teachers' solutions and common mistakes. A control group had only notes and the textbook.

• During practice, with AI available, scores rose 48% for GPT Base and 127% for GPT Tutor compared with the control group.
• On the exam that followed, taken without AI, the GPT Base group scored 17% lower than the control group.
• The GPT Tutor group scored about the same as the control group: no harm, but no gain either.

The authors found that students used the unrestricted tool as a "crutch", most often by asking for the answer and copying it. Students in the GPT Base group also did not believe they had learned less.

The Harvard physics study (Kestin et al., Scientific Reports)

In a 2025 randomized controlled trial in Scientific Reports, Greg Kestin and colleagues compared a custom GPT-4 tutor with in-class active learning in Harvard's introductory physics course for life sciences students (194 students). The tutor was designed around teaching best practices: active engagement, managing cognitive load, a growth mindset, and moving through each part of a problem in order, with instructor-written solutions behind the scenes.

• Median learning gains in the AI group were more than double those of the in-class group.
• The median student spent 49 minutes with the AI tutor, compared with an estimated 60 minutes of learning time in class.
• Students reported feeling more engaged and more motivated with the AI tutor.

The takeaway for students: the same model can help you learn or help you skip learning. The design, and how you use it, decides which.

Which AI tutors work this way, and who can use them?

Several tools follow the tutor-not-solver approach, but access varies: some are open to anyone, while others are campus pilots tied to one university or course.

Each is useful for something. Study mode is the easiest to try if you already use ChatGPT, though OpenAI notes its custom instructions can behave inconsistently. Khanmigo also offers writing coaching and code review. PeteChat and the IU coach show where campus tools are heading: tied to your real course materials and rules, but only if your school builds one.

What should you look for in an AI tutor?

Look for a tool that asks for your materials, hints before it answers, checks you understood, respects your course rules, keeps your work yours and is honest about its limits. Use this table to judge any tool.

The last row matters. In the PNAS study, the unrestricted GPT-4 tool answered the practice problems correctly only 51% of the time. A tutor that shows its steps lets you catch errors; a copied answer does not.

How can you make any chatbot act more like a tutor?

Tell it the rules at the start and give it your real materials. A general chatbot can follow tutor instructions, but a prompt is a request, not a lock: if you ask for the answer, it may simply give it.

1. Check your course rules. Read the syllabus and the assignment instructions before you start.
2. Share your materials. Paste the question, the relevant lecture notes and the rubric if there is one.
3. Set the tutor rules. Start the chat with the prompt below.
4. Attempt every step yourself. Type it even if you think it's wrong; that is what the tutor reacts to.
5. Finish with a cold check. Ask for two similar questions with no hints and answer them without help.

If you use ChatGPT, study mode does much of this for you. Still, guardrails built into a tool are sturdier than a prompt you can drop when frustrated.

Is using an AI tutor allowed in your course?

It depends on your school, and often on the individual course or assignment. Check before you start, not after.

For example, the University of Toronto Mississauga's academic integrity guidance says students may not use generative AI in a course unless the instructor explicitly permits it. It also warns that rules can differ between courses and assessments, and tells students to ask their instructor whenever they are unsure. Other universities set different defaults.

The confusion is common: in a small survey of 20 Indiana University students, the AI Learning Coach team found 90% had avoided AI for schoolwork for fear of violating a policy. Check the syllabus, then the assignment instructions, then ask.

Where CMPUS fits

CMPUS is a guardrailed AI study platform that any university student can use directly, without waiting for their school to run a pilot. Instead of one general chatbot, it splits the work across specialist AI "Classmates", each built for one academic job and none designed to hand back finished assignments.

• Alex, the front desk, routes you to the right specialist.
• Leo, the Personal Tutor, explains at your level and checks it landed.
• Quinn, the Mock Tester, builds practice tests from your own notes, scores answers and flags weak spots.
• Jordan gives a practice grade with criterion-by-criterion feedback against your rubric.
• Maya plans essays with you (thesis, outline, source plan); you write them.

Each Classmate asks for what it needs first, such as your brief, rubric or notes, so you don't need to write tutor prompts. Try it free as a Guest (5 interactions a day, no sign-up) or see how it differs from ChatGPT. Your course's AI rules still apply.

Try it on CMPUS

Leo explains at your level, then checks that it actually landed before moving on.

LeoPersonal TutorQuinnMock Tester

Enter the CMPUS 5 free interactions a day. No sign-up, no card.

Questions people ask

Sources

1. Generative AI without guardrails can harm learning: Evidence from high school mathematics · PNAS
2. AI tutoring outperforms in-class active learning: an RCT introducing a novel research-based design in an authentic educational setting · Scientific Reports (via PubMed Central)
3. Introducing study mode · OpenAI
4. Khanmigo for learners · Khan Academy
5. Tutor, Not Solver: Designing a Guardrailed AI Assistant for Learning in Higher Education: A Design Case of PeteChat · arXiv (Purdue University authors)
6. AI Learning Coach · Indiana University GenAI Project Showcase
7. GenAI Project Showcase · Indiana University UITS Student Outreach
8. Generative AI and Academic Integrity · University of Toronto Mississauga

CMPUS guides are general advice, not your course's rules. Your syllabus and your instructor decide what is allowed. Spotted something out of date? Tell us.