How to summarise lecture notes with AI, and then actually learn them

When is an AI summary actually useful?

Summarising is one of the most common things students ask AI to do. In HEPI's 2025 survey of 1,041 UK undergraduates, using AI to summarise articles had grown more than any other use and was the second most popular, behind explaining concepts (58%).

It is worth being clear about what a summary can and cannot do for you. In their large review of study techniques, Dunlosky and colleagues (2013) rated summarisation as low utility, alongside highlighting and rereading. Summarising helped some students on some tasks, but only under limited conditions. And that review was about students writing their own summaries. When an AI writes it, you skip even the effort of deciding what matters. Tilburg University's library guide on AI and note-taking makes the same point: an AI summary of your notes has limited value because you miss the work of condensing the material yourself.

So a summary is useful as a starting point:

• Before a lecture or a first read, to see the shape of a topic.
• When you have fallen behind, to decide which parts of a lecture need your time first.
• As a skeleton for questions, which is where the learning actually happens (see below).

It is not useful as the thing you read over and over before the exam.

How do you get a good summary from ChatGPT, Claude, Gemini or NotebookLM?

The general steps are the same in all of them.

1. Give it the file, not the topic. Attach the lecture slides or PDF (ChatGPT, Claude and Gemini all accept file uploads), or add it as a source in a NotebookLM notebook. If you only type "summarise the lecture on the cardiac cycle", the tool answers from everything it has seen, not from what your lecturer said.
2. One lecture at a time. A whole module in one go gives you a summary too thin to be useful and too long to check.
3. Check the text came through. Slides that are mostly images or scans may not come through as text. Ask it to quote the title of slide 5. If it cannot, the summary will be built on guesswork.
4. Ask for structure, not prose. A wall of paragraphs is hard to check and hard to turn into questions. Headings that match the lecture's own sections are easier on both counts.

If you are using ChatGPT with a PDF, our step-by-step guide to uploading a PDF to ChatGPT to study covers the upload itself and the prompts that keep it inside your file.

What should the prompt say?

Here is a prompt that works across tools. Paste it with the file attached:

"Summarise the attached lecture using only this file. Follow the lecture's own order and use its section headings. Under each heading, give the main points in short bullets. Define each key term using the wording from the file, in quotation marks. After each bullet, give the slide or page number it comes from. Keep every qualifier the lecture uses, such as 'in some patients' or 'may'. If something is unclear or missing in the file, say so rather than filling it in."

Each part does a job:

• "Using only this file" narrows what it draws on. It is an instruction, not a guarantee.
• "The lecture's own headings" keeps the summary lined up with how your course is organised, which makes checking much faster.
• "In quotation marks" for definitions stops it paraphrasing the exact wording you may need in an exam.
• Slide or page numbers turn every bullet into something you can check in seconds.
• "Keep every qualifier" targets the most common kind of summary error, below.

How do you check the summary against the source?

Assume the summary contains mistakes until you have looked. The research on AI summaries gives you a good idea of what kind.

Peters and Chin-Yee (2025) collected 4,900 summaries of scientific texts written by ten large language models. Most of the models tended to state conclusions more broadly than the original studies did, dropping the hedges and limits. Compared with human-written summaries of the same articles, the AI summaries were nearly five times more likely to contain broad generalisations. Asking the models to be accurate did not fix it; in the press release the authors report that it made overgeneralisation nearly twice as likely.

A separate study coordinated by the European Broadcasting Union and led by the BBC (2025) had journalists check more than 3,000 answers from four AI assistants about news stories. 45% had at least one significant issue, 31% had serious sourcing problems and 20% had major accuracy issues such as invented or outdated details. That study tested questions about the news rather than summaries of one file you supplied, but the failure types are the ones to look for in your own summary.

A quick check, one heading at a time:

• Qualifiers: did "can", "may", "in some cases" or "in adults" survive? A summary that turns "may increase" into "increases" is wrong in exactly the way exams test.
• Numbers, names and dates: compare each one with the slide it cites.
• Definitions: search the file for a few words of each quoted definition. If you cannot find it, it was not quoted.
• Missing sections: skim the slide titles. A summary can quietly skip the hardest part of a lecture.

Our guide to checking AI study answers goes further on spotting confident errors.

How do you turn a summary into revision?

This is the step that makes the summary worth having. Reading it again is still reading. Testing yourself on it is retrieval practice, which Dunlosky and colleagues rated as one of only two high-utility techniques, alongside spreading your study out over time.

1. Turn each heading into a question. "Phases of the cardiac cycle" becomes "What are the phases of the cardiac cycle, in order, and what happens in each?"
2. Cover the bullets and answer from memory. Out loud or on paper. Then check, and mark what you missed.
3. Ask the AI to quiz you on the same file. "Using only the attached lecture, ask me one question at a time on the sections I got wrong, and wait for my answer." Our guide on using ChatGPT to revise has more prompts for this part.
4. Come back to the misses. Ask them again tomorrow and in a few days, not just tonight.

The summary told you what was in the lecture. The questions tell you what is in your head.

If you would rather skip from file to questions, Ace turns the lecture file itself, PDF, slides or Word document, into practice questions written only from that material. Every answer is checked against the file, and the sentence that backs it is shown under each answer, so the check against the source is built in. See how it works.

Frequently asked questions

Sources

1. Dunlosky, J., Rawson, K. A., Marsh, E. J., Nathan, M. J., & Willingham, D. T. (2013). Improving students' learning with effective learning techniques: Promising directions from cognitive and educational psychology. Psychological Science in the Public Interest, 14(1), 4–58.
2. Peters, U., & Chin-Yee, B. (2025). Generalization bias in large language model summarization of scientific research. Royal Society Open Science, 12(4), 241776.
3. European Broadcasting Union & BBC (2025). Largest study of its kind shows AI assistants misrepresent news content 45% of the time, regardless of language or territory. EBU press release, 22 October 2025.
4. Freeman, J. (2025). Student Generative AI Survey 2025 (HEPI Policy Note 61). Higher Education Policy Institute.
5. Tilburg University Library. Note-taking: AI use. Tilburg University LibGuides.