Read the source, use the summary as a map

Use it when a paper, report or long article will feed a decision, and an assistant offers to summarise it.

Not when a skim is enough. For low-stakes answers, Check before you trust is the five-minute version.

You need the full text of the source in a form you can search, about 45 minutes, a place for notes, and an assistant that accepts the file or pasted text. Paste only what you are allowed to paste: What not to paste. The steps, times and prompts are our suggestion, and no study tests them.

Steps

The prompts

For step 2, paste this into a new chat with the source attached or pasted:

I have attached <SOURCE_TITLE>. Do not summarise it. Make me a map:
1. List its sections or headings in order, with one line on what each covers.
2. Quote the exact sentence that states the main finding or claim, and say
   where it is (page or section).
3. Quote the exact sentence or passage where the authors state the
   limitations, and say where it is.
4. Say which sections bear on this question: <MY_QUESTION>
If you cannot find a sentence, say so. Do not paraphrase inside quotation marks.

Replace the following:

• <SOURCE_TITLE>: the title of the paper or report.
• <MY_QUESTION>: the sentence from step 1.

For step 7, paste this into the same chat as the map, so the source is still attached:

Below are five bullets I wrote from memory after reading parts of
<SOURCE_TITLE>. For each bullet:
- quote the sentence in the source that supports it, with its location;
- say who or what was studied (who, how many, where);
- say whether my bullet is broader than, the same as, or narrower than what
  the source supports;
- if no sentence supports it, mark it UNSUPPORTED.
Do not add new points.
My bullets:
<MY_BULLETS>

Replace the following:

• <SOURCE_TITLE>: as above.
• <MY_BULLETS>: your five bullets, pasted as a list. Leave out the doubt.

Then search the source for each supporting sentence, as in the Check above.

How to tell it worked

• You can say, without looking, what the main finding is, who was studied, and what the authors say they could not show.
• At least one bullet changed after step 7, or none did and you read the limitations in full.
• It has failed if your bullets match the assistant's map word for word. Go back to step 5 and read.

Variations

• A book chapter. In place of steps 6 and 7, ask: "Ask me five questions about <CHAPTER>, one at a time. Wait for my answer, then say what the text says, with its location, and mark my answer right, partly right or wrong." Check each location it gives.
• A quick version. Steps 1 to 5 with only the main-finding and limitations sections, and three bullets in step 6. About 30 minutes.

Why read the original

• In a study of 4,900 summaries from 10 leading models, three of the models overgeneralised scientific findings in 26 to 73% of cases. In a smaller test of three of the models on 100 medical articles, their summaries were nearly five times as likely as expert-written summaries to generalise broadly (odds ratio 4.85, 95% CI 3.06 to 7.70). Asking for accuracy did not help: most models generalised more, and newer models tended to do worse. Claude models did not differ significantly from the originals (Peters & Chin-Yee, benchmark, peer-reviewed). The models tested were released by early 2025, so we make no claim about current models. That is why step 7 checks scope.
• In seven experiments with 10,462 participants, people who learned a topic from an AI summary rather than web links developed shallower knowledge, even with the same facts and live links (only 26% clicked them). The advice they then wrote was sparser and less original, and recipients found it less persuasive (Melumad & Yun, PNAS Nexus, Oct 2025, randomised experiments). That is why you read the original.

Sources

1. Peters & Chin-Yee. Overgeneralisation in AI summaries of scientific findings, Royal Society Open Science, 2025. Peer-reviewed benchmark. Link
2. Melumad & Yun. Learning from AI summaries compared with web links, PNAS Nexus, 2025. Randomised experiments. Link

Changes

1. v1.010 Oct 2026first version