Summarize a long document

~ min read

30-second summary

“Summarize it” on its own gives you the Wikipedia version: correct, but not tailored to you.
Three techniques: say how you want it (length, form, depth), say why you’re reading it (the goal), map first, details later (for long documents).
On a long document the AI can drop important parts or invent new ones. That’s the specific risk of this lesson.
Two practical countermoves: ask it to quote the exact sentence, and do a spot-check by reading a section yourself in the original document.
PDF, paste, link: each AI handles files a bit differently. Knowing what yours accepts saves surprises.

You saved an 8,000-word article for “later”, you got a 40-page PDF in a work chat, you started a book and want to hold on to what it says. Three different situations, the same problem: little time, a desire to shrink that text down to something you can read in two minutes.

Summarizing is one of the AI’s most natural use cases. It’s good at it, and you pick it up fast. But it has a specific trap: on a long document, the AI can hand you a perfectly structured summary that leaves out an important piece, or worse, invents a claim that isn’t in the text. This lesson gives you the techniques to ask well and, above all, to verify.

Compared to Understanding something complicated the scale changes. There the document was hard but short (a bill, a clause, a few pages of rules); here it can be easy but long, or long and hard at the same time. A 30-page contract is both: in that case the techniques from the two lessons combine. First you make the language clear, then you extract what matters to you. Verifying, on a long document, costs more and works differently.

The three techniques

Say how you want it

“Summarize it” on its own almost always produces the same thing: a generic, encyclopedia-style paragraph. Without instructions the AI assumes an average reader and gives you the summary that “works for anyone”, which is different from “works for you”. To get something useful you need at least two pieces of guidance out of three:

How long: “in ten bullet points”, “in 200 words”, “in three paragraphs”.
In what form: “bulleted list”, “narrative paragraph”, “a two-column table, topic and main conclusion”.
How deep: “just the main concepts” if all you need is the gist, “examples included” if you’re using it to study, “all the nuance” if you’re going to discuss it with someone.

“Summarize this article in ten bullet points, one per main topic, with one line of detail for each.”

It’s a precise prompt: reuse it on different documents and you get summaries in the same shape, comparable with each other. If you don’t like the answer, you change one piece in the next turn (“too long, give me five points”) without rewriting from scratch. It’s the mechanism of Iterate the conversation.

Say why you’re reading it

This is the move that makes the difference, and most people don’t use it. Instead of asking for a generic summary, you tell the AI the goal you’re reading the document for.

“Summarize this report so I can decide whether it’s worth reading in full.”

“Summarize this article for a colleague who has to attend a meeting about it but doesn’t have time to read it.”

“From this chapter give me only what a high school teacher needs to explain the topic in class.”

The AI doesn’t just compress: it filters. It changes what goes in the foreground and what gets left out. The same document, with three different goals, produces three different summaries. It’s the same principle as Ask well: say the outcome you want, not just the topic.

Map first, details later

For really long texts (an academic paper, a book, a 50-page annual report), the monolithic summary is the pattern that holds up worst: the AI has to compress too much and the details that mattered to you evaporate.

Better to do it in two passes. First: ask for the map of topics.

“Break this document into sections. For each, one line on what it’s about, without going into the content.”

You’ll get something like: “section 1, introduction and background; section 2, research method; section 3, main findings; section 4, study limitations; section 5, recommendations”. From that map you can see where the value is for you.

Second: you zoom in only on what you need.

“Give me a detailed summary of section 3 and section 5. Leave the others as they are, I’m not interested in them right now.”

It’s the same way you’d read the book in practice: first the table of contents, then you zoom where it counts. A good map can also tell you, sometimes, that the document isn’t for you at all, and spare you an hour of reading.

It also works on documents without pre-titled sections: a long newspaper article, an essay, a chapter of fiction. The AI builds the map for you, identifying topic shifts and proposing a “derived” index. The difference with Technique 1: there you ask for a summary of the whole document in one shot; here you break the work into two passes, and you also get a compass to orient yourself.

Three routes, in order of reliability. Upload: almost every AI accepts PDF, Word, plain text. In the chat, look for a paperclip or “plus” icon near the text box, select the file, then type your question. Copy and paste: great for web articles and not-too-long documents, as long as the text fits in a single message. Link: some AIs (ChatGPT, Gemini, Perplexity) open the page and read it; others (Claude, by default) don’t browse the web and will ask you to paste the text.

One edge case: on a whole book or documents hundreds of pages long, every AI has a threshold above which the document is “too big”. Sometimes it tells you with an error and won’t let you upload. Sometimes the file goes through but the AI only reads part of it and answers as if it had read everything, without warning you. When that happens, split the document into chunks (chapters, blocks of pages) and feed it one chunk at a time, or switch to a dedicated tool like Google NotebookLM, free and built for analyzing very long documents.

An example: a 40-page report

Let’s try “map first, details later” on a realistic case. You’ve just uploaded into the chat an OECD report on hybrid work after the pandemic, 40 pages, and you want to figure out whether it’s worth reading in full.

You

I've uploaded an OECD report on post-pandemic hybrid work, 40 pages. Before I read it in full, break it into sections and give me one line for each, without going into detail.

The report is laid out in six sections. 1. Opening summary: the three key messages of the report, in short form. 2. Context: how hybrid work spread across OECD countries from 2020 to 2024. 3. Data and methodology: which surveys and sources were used, and their limitations. 4. Effects on the labor market: productivity, wages, geographic mobility. 5. Effects on workers: well-being, mental health, gender differences. 6. Policy recommendations: suggestions to member-country governments. Each section is 5-8 pages, except the opening summary (2 pages) and the recommendations (4). The report is laid out in six sections. 1. Opening summary: the three key messages of the report, in short form. 2. Context: how hybrid work spread across OECD countries from 2020 to 2024. 3. Data and methodology: which surveys and sources were used, and their limitations. 4. Effects on the labor market: productivity, wages, geographic mobility. 5. Effects on workers: well-being, mental health, gender differences. 6. Policy recommendations: suggestions to member-country governments. Each section is 5-8 pages, except the opening summary (2 pages) and the recommendations (4).

From the map you can tell right away: if you care about the well-being side, section 5 is the one you shouldn’t skip. Second pass, focused.

You

Give me an extended summary of section 5, around 300 words. Leave the other sections as they are.

Section 5: Effects on workers. The report splits the effects into three groups. First, mental health. Workers in a hybrid arrangement report lower stress levels than people fully on-site, but more sense of isolation. The net effect depends on the dose: two or three days from home a week give the best balance; less or more makes it worse. Second, gender differences. Women with small children benefit the most from hybrid work in terms of labor-force participation. At the same time they're the most exposed to domestic overload, because physical presence at home tends to redistribute care work onto them, not onto the partner. Third, careers. The report highlights a wage and promotion penalty for those who work remote more than three days a week, especially in junior roles. The mechanism: less visibility to the manager, fewer informal contacts with people who could help you grow. The report closes by noting that the three effects don't cancel out, and suggests governments treat them separately rather than as a single net balance. Note: in the report these points are backed by tables and charts not reproduced in the text; if you need the exact figures, it's worth opening pages 22-28 of the original. Section 5: Effects on workers. The report splits the effects into three groups. First, mental health. Workers in a hybrid arrangement report lower stress levels than people fully on-site, but more sense of isolation. The net effect depends on the dose: two or three days from home a week give the best balance; less or more makes it worse. Second, gender differences. Women with small children benefit the most from hybrid work in terms of labor-force participation. At the same time they're the most exposed to domestic overload, because physical presence at home tends to redistribute care work onto them, not onto the partner. Third, careers. The report highlights a wage and promotion penalty for those who work remote more than three days a week, especially in junior roles. The mechanism: less visibility to the manager, fewer informal contacts with people who could help you grow. The report closes by noting that the three effects don't cancel out, and suggests governments treat them separately rather than as a single net balance. Note: in the report these points are backed by tables and charts not reproduced in the text; if you need the exact figures, it's worth opening pages 22-28 of the original.

In two passes you got a map of the whole document and a detailed summary of the section you needed. If that’s enough, stop here. If you want to go deeper, one more pass.

What to always verify

On long documents the AI gets things wrong in a specific, slightly sneaky way: it presents as present in the text claims that aren’t actually there. It’s not malice, it’s the mechanism you saw in When to trust it (and when not): the AI produces text that sounds true, and a summary of a document can contain a plausible statement that the document doesn’t actually make.

It happens more on long documents for two reasons. One: the document can exceed the AI’s capacity, and the AI then reads only part of it and fills in the gaps with plausible-but-unverified claims. Two: on a three-page contract you’d see right away that the sentence isn’t there; on a 50-page report, who’s going to go check?

Two simple countermoves.

Ask it to quote the exact sentence. If it says “the report argues that hybrid work reduces stress”, reply: “quote the sentence of the report you got this from, and tell me which page it’s on”. If it quotes, go check it in the original document. If it can’t point to it, or it gives you a page number but the sentence isn’t on that page, you’ve found an invention. This check works even when the topic is new to you: you don’t need to be an expert, you just need to compare what the AI says with what’s written.

Do a spot-check yourself. Take one of its claims and go read that page in the original document. You don’t need to check everything, one is enough: if that one’s correct, the rest of the summary is probably reliable; if it’s wrong or inflated, the suspicion extends to the whole output.

What comes next

Summarizing is the base form. When you want that document to become material for learning something, the techniques change: you ask open questions, you ask for explanations at different depths, you let the AI quiz you back to see what you haven’t internalized. The next lesson, Learn something new, is dedicated to this.

Check what you understood

About Privacy Open manual, written by a real person, in collaboration with AI.