QWE AI Academy Two students. Same biology midterm. Both use ChatGPT to make study notes.
Student A types: “Make me study notes on cellular respiration.” Student B uploads her lecture slides, the assigned textbook chapter PDF, and last week’s problem set, then asks ChatGPT to build notes only from those files. Student B’s notes match what’s actually on the exam. Student A’s notes contain a citation to a paper that doesn’t exist and a definition her professor would mark wrong.
That’s not a hypothetical. It’s the central problem of using ChatGPT to create study notes, and almost every tutorial online glosses over it.
Why the standard “summarize this topic” approach falls apart
The default move – ask ChatGPT for notes on a topic – pulls from its training data. Which sometimes is great. And sometimes it confidently invents things.
How often? According to a 2024 Sage paper by Buchanan, Hill, and Shapoval, GPT-3.5 produces false citations at a rate over 30%, and GPT-4 still comes in above 20%. The kicker is buried in the same study: the narrower the prompt, the higher the false-citation rate – and niche course content is about as narrow as it gets. Translation: the more specialized your course material, the more ChatGPT makes up.
A separate JMIR study from 2024 found hallucinated references exceeded a 25% rate when LLMs were asked to support systematic reviews.
If you’re studying “photosynthesis” or “the French Revolution,” topic prompts are mostly fine. If you’re studying “the second-quarter content from your Comparative Politics professor’s lectures,” they’re a trap.
The upload-first workflow (what actually works)
The fix is structural: don’t let ChatGPT generate from memory. Force it to work only from materials you trust. Here’s the sequence.
- Collect the source material. Lecture slides (PDF), your raw class notes, the assigned reading, and any handouts. ChatGPT accepts PDFs, images of handwritten notes, and pasted text.
- Turn on Study Mode. According to OpenAI’s announcement, select Tools in the prompt window and choose Study and learn from the drop-down menu, or go to chatgpt.com/studymode.
- Upload, then constrain. Attach your files in one message and write: “Build study notes only from the attached files. If something isn’t covered in the files, say ‘not in source’ instead of filling it in.” That last sentence is the whole game.
- Ask for structure, not summary. Request H2 topics, H3 sub-concepts, three to five bullets each, plus a glossary of terms that appear in your sources.
- Generate retrieval prompts. Once notes are clean, ask for cloze-deletion flashcards and short-answer questions from your notes. Spacing and retrieval practice are two of the most reliable strategies for long-term retention (Nature Reviews Psychology, 2022) – so the flashcards aren’t decoration. They’re the point.
The constraint in step 3 is what most guides skip. Without it, ChatGPT silently “helps” by adding context from its training data, which is exactly where the fabrications live.
What ChatGPT Study Mode actually does
On July 29, 2025, OpenAI announced Study Mode – a tutor-style feature that holds back direct answers in favor of step-by-step questioning. Available to logged-in users on Free, Plus, Pro, and Team at launch, with ChatGPT Edu rolling out in the weeks after.
Under the hood it’s not a different model. Study Mode runs on custom system instructions – OpenAI’s stated reason being that they wanted to learn from real student feedback before baking behavior into the model itself. The side effect: behavior can be inconsistent across conversations.
For note creation, the two behaviors that matter most: it works with PDFs and images you upload, and it asks Socratic follow-up questions that surface gaps you didn’t know you had. The rest – breaking concepts into sections, memory personalization, open-ended checks – are useful but secondary. The PDF support is what makes the upload-first workflow possible.
Pro tip: Study Mode is a toggle, not a model. Turn it off when you want raw drafting (“reformat these bullets as a one-page cheat sheet”) and on when you want active recall (“quiz me on these notes one question at a time, don’t reveal answers”). Same chat, different jobs.
A real example: studying for an econ midterm
Last semester I tested this with a friend’s intermediate microeconomics course. She had 60 pages of lecture slides on consumer choice theory and a habit of making notes that were 80% transcription, 20% understanding.
We uploaded all six lecture PDFs. First prompt:
Source: attached PDFs only.
Task: Build study notes for an undergraduate intermediate micro midterm.
Format: H2 per lecture, H3 per major concept, 3-5 bullets each.
For every formula, include: definition, when it's used, one common mistake.
If something isn't in the PDFs, write [not in source] - do not fill it in.
End with a glossary of 10-15 terms that appear in the slides.The output was 2,400 words, properly hierarchical, and – this matters – flagged seven items as [not in source] that the slides genuinely didn’t cover. Those flags became her list of professor-office-hours questions.
Then we flipped Study Mode on and pasted the notes back: “Quiz me on these notes. Multiple choice. One question at a time. After each, tell me if I’m right and ask a follow-up ‘why’ question.” The Socratic follow-ups caught two genuine misunderstandings she didn’t know she had.
The gotchas nobody mentions
Study Mode has cracks. Here’s what to watch before finals week.
| Gotcha | What happens | Workaround |
|---|---|---|
| Hallucinations scale with specificity | The narrower your topic, the more ChatGPT invents – above 20% false-citation rate for GPT-4 | Always upload sources; require [not in source] tags |
| Study Mode is bypassable | Students can switch it off and get a direct answer. No guardrails currently exist, per OpenAI. One early user reported it caved and did her writing for her after she pushed back | Self-discipline. There’s no technical fix yet |
| Memory cross-contamination | If Memory is on, biology notes from week 4 can leak into unrelated chats months later | Use Temporary Chat or turn off Memory for course work |
| Data retention | Study Mode collects the same conversation data as regular ChatGPT – retained and potentially used for model training unless you opt out (Linewize, citing OpenAI Terms of Use) | Disable “Improve the model for everyone” in Data Controls before uploading copyrighted course PDFs |
One week at a time
Treat ChatGPT as a notes processor, not a notes source. The raw material is your professor’s slides, your readings, your scribbles. ChatGPT cleans, structures, and quizzes. That’s its job here – and it’s a narrow one on purpose.
Open this week’s lecture. Upload the slides. Try the prompt from the econ example above – and see how many [not in source] flags come back. Those are the questions worth bringing to your TA.
FAQ
Is ChatGPT Study Mode free?
Yes. Available to logged-in users on Free, Plus, Pro, and Team at launch, with Edu rolling out shortly after. No paid tier required.
Can I trust ChatGPT’s notes for a high-stakes exam like the bar or MCAT?
Not without verification. False citation rates above 20% for GPT-4 – and climbing as topics narrow – make raw topic prompts genuinely risky for high-stakes prep. The safer path: use it to process your authorized materials (textbooks, official prep books, past papers) into structured notes and flashcards. Don’t ask it to teach a topic from scratch and treat the output as exam-ready.
What’s the difference between ChatGPT Study Mode and Claude or Gemini’s versions?
Same basic idea, different implementations. Anthropic’s version for Claude is called Learning Mode (launched April 2025); Google has tested “Guided Learning” for Gemini. All three default to Socratic questioning. The real differentiator isn’t the tutoring style – it’s how each handles files and whether it honestly flags what it doesn’t know. Run the same upload-first prompt in two of them and see which one returns more [not in source] tags. That one is doing the job correctly.