Can you use AI to summarise a PDF?

Yes, and for most people most of the time it works. The catch is what’s inside the PDF. A published report, a product manual or a research paper is one thing. A contract with your name on it, a medical letter or a work file full of other people’s details is another. So the verdict is yes, with care, and the care mostly comes down to whose document it is and who else might end up seeing it.

Here’s the bit people skip. When you upload a file to a chatbot, it isn’t handed to a machine and forgotten. Take Google’s Gemini as an example, since its privacy hub spells things out. The information you provide includes your prompts and your uploaded content, files included. Some of that data is looked at by human reviewers, including trained reviewers from service providers. Google’s own advice is plain: don’t enter confidential information that you wouldn’t want a reviewer to see, or that you wouldn’t want Google to use to improve its services.

How to do it well

Start with the right kind of document. Public material is fine, and so are your own non-sensitive files. If it’s a published report or a manual you’d happily leave on a café table, you’re not risking much by pasting it into a summariser.

Then think about how you ask. “Summarise this” gets you a tidy paragraph that sounds confident, and that’s the trouble. Ask for structure instead: key points, numbers, dates and decisions, with page references so you can check them. Page numbers turn a vague summary into something you can verify in a minute or two.

Summarise this document in ten bullet points with page numbers. List any deadlines, amounts and obligations separately. If something is missing, say "not stated".

That last line matters more than it looks. Telling the tool to say “not stated” gives it a way out other than filling the gap with something plausible. It helps; it doesn’t make the tool immune.

You can also ask it to quote the passage behind each key claim, then spot-check a handful against the PDF itself. You’re sampling, like tasting soup before you serve it. If the quotes line up with the page, you can relax a bit. If one of them doesn’t exist, you’ve learned how far to trust the rest.

Long or scanned PDFs need extra suspicion. Tools can miss text that sits inside images, and tables and footnotes are easy casualties. The figure buried in a table is often the thing you most wanted, so check the tables yourself.

And for anything with legal or financial consequences, don’t lean on a summary alone. Read the relevant sections, or get advice. A summary is a map, and you wouldn’t navigate a mortgage with only a map. If you’re summarising for study, can you use AI to help me study? has more.

Rules and risks

Privacy of uploads. The UK’s National Cyber Security Centre (NCSC) is direct about this. Queries to public AI chatbots are visible to the provider, they’re stored, and they will almost certainly be used for developing the service or model at some point. Providers or their partners and contractors may be able to read them. The NCSC recommends keeping sensitive information out of queries to public tools, not submitting anything that would cause problems if it were made public, and reading the terms of use and privacy policy before you start. Private or enterprise arrangements are a different thing.

Human review and retention. This one surprises people. With Gemini, chats that have been reviewed by humans aren’t deleted when you delete your activity; they’re kept for up to three years. Temporary chats aren’t used to train Google’s AI models. They’re kept with your account for 72 hours, as are chats with Keep activity off, and Google still uses them to respond and to protect users, including with help from human reviewers. The default auto-delete for Gemini Apps activity is 18 months; you can change it to 3 months, 36 months or indefinite, and delete chats manually.

Workplace data. If the PDF belongs to your job, stop and think. Work files, contracts, medical letters, anything containing other people’s personal data: these aren’t yours to feed into a public tool on a whim. Check your employer’s policy first, use a work-approved tool if there is one, or strip out names and identifying details before you upload. Where you do use a public tool, prefer temporary or no-history modes, and check the training settings.

Accuracy. The Department for Education notes that generative AI can produce inaccurate or false information and present it as fact, and that its output needs critical judgement to check for accuracy. That applies to summaries as much as anything else, which is why the page references and quoted passages above are worth asking for.

This is general information, not legal advice.