How to Summarise PDFs and YouTube Videos With AI Accurately
Short answer: To summarise a PDF or YouTube video with AI, upload the PDF (or give the video link or transcript) to NotebookLM, Gemini, ChatGPT or Claude and ask for a structured summary with a page number or timestamp for every point. Then check the points that matter against the source, because AI summaries most often drop exceptions, conditions and figures.
Getting a summary is easy. Getting one you can rely on is the hard part. A neat ten-bullet summary of a 60-page document reads as complete whether or not it skipped the one clause that affects you.
This guide is for one-off jobs: a policy PDF this week, a lecture video tonight, in whichever assistant you already use. It shows which tool handles which source, a simple accuracy test you can run yourself on any 60-page PDF and 40-minute lecture, the kinds of omissions to look for, and prompts that make every point traceable to a page or timestamp. If you will study the same set of files all term, a NotebookLM notebook is the better setup; our NotebookLM study guide covers that.
Which AI tool can summarise PDFs and YouTube videos?
As of November 2026, all four major assistants summarise PDFs; they differ more on video links and on how clearly they show where each point came from.
| Tool | PDFs | YouTube videos | Shows where points came from |
|---|---|---|---|
| NotebookLM | Yes, several at once | Yes, by link for public videos with captions | Yes, clickable citations to the source passage |
| Gemini | Yes | Yes, by link in many cases | Partly; ask for timestamps |
| ChatGPT | Yes | Not reliably by link; paste the transcript | Only if you ask for page numbers |
| Claude | Yes, strong on long files | Paste the transcript | Only if you ask for page numbers |
For study material you will return to, NotebookLM's clickable citations make checking fastest, which is why it gets its own notebook-based guide linked above.
How to run your own accuracy test (60-page PDF, 40-minute lecture)
The only way to know how far to trust a summariser on your kind of material is to test it on a document you know. This takes about an hour once and saves guessing afterwards.
- Pick a document you have read properly, around 60 pages: say an HR policy handbook, an insurance policy wording or a government scheme guideline.
- Before using any AI, write your own list of the 10 points a summary must not miss. Include at least two exceptions or exclusions, two figures (amounts, limits, dates) and one point from the last quarter of the document.
- Run the same prompt (below) in three tools, for example NotebookLM, Gemini and ChatGPT, each in a fresh chat.
- Mark each of your 10 points as captured, partly captured (vague or missing the number) or missed, and note any statement the summary makes that is not in the document.
- Repeat with a 40-minute lecture video: list 10 must-catch points with their timestamps first, then compare.
- Keep the scorecard. It tells you which tool to use for which kind of source, and how much checking you need.
Scorecard template for marking omissions
Use a simple table like this. The rows show the kinds of points worth testing; fill the tool columns with your own results.
| Must-catch point | Type | Where in source | Tool A | Tool B | Tool C |
|---|---|---|---|---|---|
| Main purpose of the document | Overview | First pages | Captured / partial / missed | ||
| Key exclusion or 'does not apply if' clause | Exception | Middle section | |||
| A money limit or percentage | Figure | A table | |||
| A deadline or notice period | Date | Any | |||
| A condition hidden in a footnote | Condition | Footnote or annexure | |||
| A point from the final quarter | Late content | Last 15 pages | |||
| Speaker's caveat in a lecture | Exception | Q&A near the end | |||
| Anything the summary added that is not in the source | Invented | Not found |
The pattern to watch for: in our experience, AI summaries are generally good at the main purpose and headline points, and weaker on exactly the rows that carry risk: exceptions, figures inside tables, footnotes, and content near the end of long sources. Your own test will show how your chosen tools behave on your material.
Prompts that force page and timestamp references
The single most useful change is asking for a reference with every point. It makes the summary checkable in seconds and tends to reduce invented content, because the model has to anchor each claim to a place in the source.
PDF prompt: 'Summarise this document for an employee who needs to know their rights and obligations. Use headings that follow the document's own sections. For every bullet, add the page number in brackets. Quote exact wording for any amounts, limits, deadlines and exclusions. End with a section called Exceptions and conditions listing every exclusion, condition or does-not-apply clause, with page numbers. If something is unclear in the document, say so rather than interpreting it.'
Video prompt (with a transcript pasted or a link in a tool that supports it): 'Summarise this lecture as a set of study notes. Give a timestamp for every point. Separate what the speaker states as fact from examples and opinions. List any caveats, corrections or answers to audience questions in their own section, with timestamps.'
Critique of both: the reader is named, so the summary prioritises what matters to them; exact quotes for figures stop rounding and paraphrase errors; a dedicated exceptions section targets the weakest spot; 'say so rather than interpreting' gives the model permission not to guess.
How do you summarise a very long PDF in parts?
For documents over roughly 80 to 100 pages, or anything dense with tables, summarise in sections rather than all at once. A single whole-document summary has to compress harder, and that is where exceptions and late content fall out.
- Ask first for the table of contents with page ranges: 'List the main sections of this document with their page ranges.' Check it against the actual PDF.
- Summarise one section at a time with the PDF prompt above, naming the page range: 'Summarise pages 21 to 34 only.'
- For any table you need, ask the tool to reproduce the table exactly before summarising it, and compare a few cells with the PDF.
- Once every section is done, ask for a one-page overview built only from the section summaries in the chat.
- Ask a final check question: 'Which parts of the document did your summaries not cover?' Read those parts yourself.
If the PDF is a scan (you cannot select the text in your PDF viewer), run it through a text-recognition step first, or expect more misread numbers. Very large files may also exceed upload limits on free plans; splitting the PDF by section solves both problems.
How do you summarise a YouTube video that the AI cannot open?
Use the transcript. Most YouTube videos have one, even if it is auto-generated.
- Open the video on a computer, expand the description and click Show transcript.
- Turn timestamps on, select all the text in the transcript panel and copy it.
- Paste it into ChatGPT or Claude with the video prompt above.
- For Hindi or Hinglish lectures, expect errors in auto-captions, especially technical terms and names. Ask the AI to list words that look like transcription errors, then check those moments in the video.
- For a series of videos on one subject, add them all as sources in one NotebookLM notebook instead.
When should you not trust an AI summary?
Do not rely on a summary alone for anything you will sign, pay, claim or comply with: insurance policies, loan agreements, tender documents, legal notices, exam rules. Use the summary to find the relevant pages, then read those pages yourself.
The same applies to figures. If a number from a summary goes into an email, a report or a decision, find it in the source first. Our guide on how to fact-check AI answers has a three-minute routine for exactly this.
For students, summaries are a revision aid, not a substitute for reading; the trade-offs are covered in free AI tools for students. Better source-based prompting starts with the habits in our guide to writing AI prompts, and summarising with checks is a core exercise in our Applied AI Professional course.
Frequently asked questions
Which AI is best for summarising PDFs?
NotebookLM is strong when you want citations back to the exact passage, and Claude handles very long files well. ChatGPT and Gemini also summarise PDFs; whichever you use, ask for page numbers with every point.
Can ChatGPT summarise a YouTube video?
Not reliably from the link alone. Copy the transcript from YouTube's Show transcript option and paste it in, or use Gemini or NotebookLM, which can often work from the link for public videos with captions.
Is there a free AI summariser?
Yes. NotebookLM, Gemini, ChatGPT and Claude all summarise documents on their free plans, with limits on file size and daily use as of November 2026. Check each tool's current limits.
Why do AI summaries miss important points?
Summaries compress, and models tend to keep headline points while dropping exceptions, figures in tables, footnotes and content near the end. Asking for page references and a separate exceptions section reduces this.
Can AI summarise a 100-page document?
Yes, most current assistants can read documents of that length. For reliable results, ask questions section by section, request page numbers, and check the key figures and exceptions yourself.
Keep reading in AI tools
- How to Write AI Prompts That Work First Time Start here
- How to Write AI Image Prompts That Match the Picture in Your Head
- Claude Artifacts Tutorial: Build Two Working Tools Without Coding
- How to use ChatGPT Projects for recurring work
- How to Use ChatGPT in Hindi and Hinglish, and Get Natural Results
Related: Applied AI Professional · How to Use NotebookLM to Study and Research From Your Own Files · All guides in AI tools