How to upload a PDF to ChatGPT to study, and keep it honest

Why upload the PDF at all?

Because your exam is on your course, not on the internet's version of the topic. If you just ask ChatGPT about "the Krebs cycle" or "the causes of the First World War", it answers from everything it has seen, which may use different terms, a different level of detail or a different emphasis from your lecturer's. Giving it your actual lecture slides or textbook chapter pulls the conversation back to what you will be examined on.

Students are already doing this at scale. In HEPI's 2026 survey of 1,054 UK undergraduates, 95% said they used AI in at least one way, and 94% said they used generative AI to help with assessed work. The question is not whether to use it, but how to use it without fooling yourself.

How to upload a PDF to ChatGPT to study, step by step

1. Pick one chapter or lecture, not the whole module. A focused file gives focused questions, and you can actually check the answers.
2. Check the PDF has real text. If you can't select the words in your PDF reader, it is probably a scan. Ask ChatGPT to quote the first line of page 2; if it can't, the text isn't coming through reliably.
3. Attach the file and give the rules first. Before you ask anything, paste one of the prompts below so the whole chat runs on those rules.
4. Answer before you look. Type your answer from memory, even when unsure. That effort is the learning.
5. Keep a list of misses. The chat won't remember your weak spots for next week. Copy each missed question into a note and ask those again in a few days.

Our guide to studying from a PDF covers the reading side; this guide is about the testing side.

Prompts that keep it inside your file

• "Use only the attached PDF. Ask me one exam-style question at a time and wait for my answer. After I answer, tell me if I'm right and quote the sentence from the PDF that proves it, with the page number. If the PDF doesn't cover something, say so instead of answering from general knowledge."
• "From pages 4 to 9 of the attached file, write a four-option multiple-choice question. Make the wrong options things a student who half-read the section would pick. Don't show the answer until I reply."
• "List the steps of the process on page 12 in a shuffled order and ask me to put them back in sequence."
• "Here are the questions I got wrong last time: [list]. Ask them again in a different form, still using only the PDF."

Two phrases matter most: "use only the attached PDF" and "quote the sentence that proves it". The first narrows what it draws on. The second makes errors visible, because a made-up answer usually comes with a quote you can't find. More prompts are in our guide to getting ChatGPT to quiz you.

How to catch it when it goes wrong

Telling ChatGPT to stick to the file helps, but it is an instruction, not a guarantee. In a 2026 study, Shoufan and Esmaeil asked 63 university students about their experience of AI "hallucinations". The problems they reported most were fabricated citations, false information, confident but misleading answers, ignoring the prompt, sticking with a wrong answer, and agreeing too readily with the student.

Simple checks catch most of this:

• Search for the quote. Use Ctrl+F (or Find on your phone) in the PDF for a few words of each quote. If it isn't there, treat the answer as unverified.
• Push back once. Tell it "I think that's wrong" when it isn't. If it instantly agrees and changes a correct answer, you have learned how far to trust its marking.
• Be extra careful with numbers, names and definitions. These are the details exams test and the ones AI most often gets slightly wrong.

Our guide to checking AI study answers goes further.

The crutch problem

The bigger risk is not a wrong answer. It is letting the AI do the thinking. In a field experiment with nearly a thousand high-school maths students, Bastani and colleagues (2025) found that access to a standard ChatGPT-style assistant improved grades on practice problems by 48%, but when the access was taken away, those students did 17% worse than students who had never had it. A version designed to give hints instead of answers largely avoided that harm.

The same pattern shows up in older research on study methods. Dunlosky and colleagues (2013) rated practice testing and spaced practice as high-utility techniques, and summarisation, highlighting and rereading as low-utility. "Summarise this PDF" is a low-utility request, however good the summary. "Test me on this PDF" is a high-utility one.

That is also the idea behind Ace: it turns your PDF, slides or lecture video into practice questions written only from that material, checks every answer against it, and shows the sentence from your file that backs each answer, so the quote check is built in. See how it works.

Frequently asked questions

Sources

1. Stephenson, R., & Armstrong, C. (2026). Student Generative Artificial Intelligence Survey 2026 (HEPI Report 199). Higher Education Policy Institute.
2. Shoufan, A., & Esmaeil, A.-A.-A. (2026). AI hallucination from students' perspective: A thematic analysis. arXiv:2602.17671.
3. Bastani, H., Bastani, O., Sungu, A., Ge, H., Kabakcı, Ö., & Mariman, R. (2025). Generative AI without guardrails can harm learning: Evidence from high school mathematics. Proceedings of the National Academy of Sciences, 122(26), e2422633122.
4. 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.