Blog ›

How to summarize a PDF with AI without getting fooled by the summary

30 September 2026 · 8 min read

Dropping a PDF into an AI chat and typing "summarize this" takes ten seconds and produces something that reads well. That is the problem: a fluent summary of the wrong things looks identical to a good one. Here is how to get a summary you can actually act on, for the three kinds of documents people summarize most, and how to check it in five minutes.

What a summary is for decides how you ask

"Summarize" is not one task. Before you type anything, answer one question: what will I do with the summary?

• Decide whether to read the whole thing → you need the topic, the conclusion, and who it's for. Five lines.
• Brief someone else → you need the points that matter to them, in the order they'd ask about them.
• Act on it (sign, pay, reply, comply) → you need the obligations, the numbers, the dates and the exceptions. Nothing else, and nothing missing.
• Learn it (a paper, a textbook chapter) → you need the argument, the evidence, and what the author would say the limitations are.

A generic "summarize this" gets you the first kind. If you needed the third, you'll be reading the whole document anyway, only later and under more pressure.

Prompts that work, by document type

These work in ChatGPT, Claude, Gemini, Copilot and most PDF chat tools. The pattern is the same each time: say who you are, say what you'll do with it, ask for a fixed structure, and ask for page references.

Contracts, leases, terms and conditions

I am the tenant/customer/supplier in this agreement. Summarize it for me
under these headings, quoting the page for each point:
1. My obligations (what I must do, by when)
2. Their obligations
3. Money: every amount, fee, penalty and how/when it changes
4. Dates and deadlines, including notice periods and automatic renewals
5. Ways out: termination, exceptions, what voids the agreement
6. Anything unusual compared with a standard agreement of this kind
Do not paraphrase amounts, dates or percentages; copy them exactly.

The last line matters. Models round, convert and "simplify" numbers when summarizing, and a "30-day notice" that was actually "3 months before the end of the term" is the kind of error that costs money.

Research papers and reports

Summarize this paper for someone who knows the field but hasn't read it:
- The question it answers, in one sentence
- The method, in two sentences, including sample size and data source
- The main result, with the actual numbers and effect sizes
- What the authors say the limitations are
- What they claim, versus what the data shows (flag any gap)
Cite page numbers. If a figure or table carries the main result, say which.

Long reports, minutes, policy documents

I have to brief my team in five minutes. From this document give me:
- The three decisions or conclusions that matter most, each with the page
- Every action item, deadline or number that someone must act on
- Open questions or disagreements the document records
- One paragraph of context for someone who has never seen it
Skip background and history unless a decision depends on it.

Where it breaks

1. Long PDFs get silently truncated

Every model has a limit on how much text it reads at once. Many tools handle a PDF that exceeds it by reading the first part and stopping, or by keeping only fragments, without telling you. The summary then covers pages 1 to 40 of a 120-page report and presents itself as complete.

Check: ask "what is on the last page?" or "list every section heading in the document". If the answer stops early, so did the reading. For big documents, summarize per chapter and then summarize the summaries, or use a tool that shows you which pages it used.

2. Scanned PDFs have no text

If the PDF is a scan (you can't select the text with your mouse), a text-based tool sees nothing, or runs OCR quietly and summarizes its misreadings. Numbers suffer most. Run OCR first with a tool that shows you the recognized text, and check the totals before you summarize.

3. Tables, footnotes and exceptions vanish

Summaries favour prose. The pricing table, the footnote that says "excluding VAT", the clause that starts "notwithstanding the above": these are exactly what a summary tends to drop, and exactly where the risk lives. Ask for them explicitly ("list every exception, exclusion and footnote").

4. The model fills gaps with plausible text

When a document doesn't say something, a summary may say it anyway, because that's what documents of this kind usually contain. A lease summary might mention a deposit that the lease never specifies. Instruct the model: "if the document does not state something, write 'not stated' rather than guessing", and treat every number you can't find on the cited page as missing.

5. Emphasis drifts toward the document's own framing

A summary inherits the author's priorities. A supplier's proposal summarized by AI reads like a supplier's proposal. If you need the other side, ask for it: "what would a sceptical buyer want to know that this document downplays?"

The five-minute check

1. Pick the three claims you'd act on (an amount, a date, a conclusion) and find each on the cited page. If a citation is wrong, assume others are too.
2. Ask for the section headings and compare with the PDF's table of contents. Missing headings mean unread pages.
3. Search the PDF for "not", "except", "unless" and "notwithstanding" and check whether the summary reflects those sentences.
4. Ask the opposite question. "What in this document argues against the conclusion in your summary?" A good summary survives it; a lazy one changes its mind.
5. Read the last two pages yourself. Signatures, annexes, effective dates and "this supersedes" clauses live there and are the most often skipped.

Which tool?

• A chat assistant (ChatGPT, Claude, Gemini) is the most flexible and answers follow-up questions. Use it for documents you're allowed to upload; check its file-size and page limits, and prefer paid tiers for long PDFs.
• Built-in readers (Acrobat's AI Assistant, Edge's Copilot sidebar, Preview's or Word's summarize features) are convenient and keep the document in place, but usually give the generic first-kind summary with less control over the prompt.
• Local models (Ollama, LM Studio with a document plugin) are the answer for confidential PDFs you can't upload anywhere. Slower and weaker on long documents; fine for a contract or a paper.
• Dedicated PDF-chat tools are worth it mainly when they show you the source passage next to each claim. That feature is the whole difference between a summary you can check and one you have to trust.