ChatGPT vs. an AI tutor: what changes when you learn with a plan
Chat assistants are great at answering the question in front of you, but learning a subject takes a plan, checks and review. Here is the difference, backed by recent studies, and a method you can use with any tool.
Short answer: a chat assistant like ChatGPT answers the question you type right now. An AI tutor works from a learning plan: it teaches in order, checks what you understood, remembers where you got stuck and decides what the next lesson should cover. To look something up, chat is enough. To actually learn a subject, you need the tutor’s structure.
What chat assistants are good at
- Looking up a definition, getting an example, fixing a piece of code, translating a paragraph.
- All of these assume you already know what to ask.
You ask, it answers, you move on. The trouble starts when you are trying to learn a whole subject: you don’t know what to ask next, earlier context drifts away as the conversation grows, and nothing checks whether you really understood.
What learning takes that chat won’t do for you
The most reliable findings in learning science all need someone to set them up:
- A plan with an order. What comes first, what comes later, and what counts as mastered.
- Recall, not rereading. In Roediger and Karpicke’s 2006 experiment, students who took repeated recall tests after reading a passage remembered 61% of it a week later. Students who only reread it remembered 40%.
- Practice spread over time. Dunlosky and colleagues (2013) rated ten common study techniques. Only practice testing and distributed practice were rated high utility; highlighting, rereading and summarizing were rated low.
- Adjusting to where you are stuck. One-on-one tutoring works because the tutor sees what you don’t understand and changes the explanation (more in Bloom’s 2 Sigma Problem).
Chat without guardrails can make you learn less
A randomized study published in PNAS in 2025 (Bastani et al.) gave nearly a thousand high school math students in Turkey one of two GPT-4 tools:
- A standard chat interface (GPT Base): practice scores went up 48%. But once the AI was taken away, exam scores were 17% lower than for students who never had it.
- A version with teaching prompts that held back answers (GPT Tutor): practice scores went up 127%, and the harm on the exam was largely mitigated.
The authors’ explanation: without guardrails, many students asked for the answer and used the AI as a crutch. They did well in practice and worse on their own.
Well-designed AI tutoring, on the other hand, can work very well. In a 2025 randomized crossover study in a Harvard physics course (Kestin et al., Scientific Reports), students who studied at home with an AI tutor built on teaching best practices had median learning gains more than double those from in-class active learning, in less time.
Put the two studies together and the difference is not AI versus no AI. It is whether there is teaching structure.
Side by side
| Chat assistant | AI tutor | |
|---|---|---|
| Starts from | One question | Your learning goal |
| How content is organized | Question and answer, following the conversation | A syllabus of things to master, taught one lesson at a time |
| Checking understanding | Only if you ask | Reflection questions every lesson, reviewed in the next one |
| Remembering you | Earlier context gets lost in long chats | Your questions, highlights and feedback are kept and carried forward |
| When you are done | Never | When the syllabus is covered, with an assessment and a summary |
How to learn this way with a plain chatbot
- Ask it to write a syllabus for your goal, listing the points you should be able to explain and use. Save it somewhere.
- Study one point at a time. Ask for 2–3 questions at the end, and tell it not to reveal answers until you have tried.
- Write your own answers first, then ask it to mark them and explain your mistakes.
- Next session, paste the previous questions and your answers back in, and ask it to review them before moving on.
- After a day, three days and a week, ask it to test you again with new examples.
This works. The catch is that you have to remember, paste and prompt every step yourself.
How Bloom does it
Bloom is a private AI tutor in your browser that builds these steps into the product:
- You write your learning goal and it drafts a syllabus of things to master. You can edit it before any teaching starts.
- Each lesson ends with 2–3 reflection questions. The next lesson opens by reviewing your answers, then covers what is still unclear.
- Select any sentence you don’t understand to ask about it. The thread stays beside that sentence and shapes the next lesson.
- When the syllabus is covered, you get an assessment and a course summary.
- You can attach your own PDFs and notes so the course is grounded in your materials.
Keep using chat for quick lookups. When you want to learn a subject properly, try a course with a plan.
References
- Bastani, H., et al. (2025). Generative AI without guardrails can harm learning: Evidence from high school mathematics. PNAS, 122(26). https://doi.org/10.1073/pnas.2422633122
- Kestin, G., Miller, K., Klales, A., Milbourne, T., & Ponti, G. (2025). AI tutoring outperforms in-class active learning. Scientific Reports, 15, 17458. https://doi.org/10.1038/s41598-025-97652-6
- Roediger, H. L., & Karpicke, J. D. (2006). Test-enhanced learning: Taking memory tests improves long-term retention. Psychological Science, 17(3), 249–255.
- Dunlosky, J., Rawson, K. A., Marsh, E. J., Nathan, M. J., & Willingham, D. T. (2013). Improving students’ learning with effective learning techniques. Psychological Science in the Public Interest, 14(1), 4–58.
Thanks for reading!