Skip to main content

How to read a research paper with AI, and check what it tells you

· 6 min read
The MindMax team
The MindMax team
Building MindMax

To read a research paper with AI without being misled, ask for the paper's central claim and what it rests on before you read anything, insist that every point comes with a verbatim quote and a page, check the two quotes that matter most, and only then read those sections properly yourself. The AI finds the way in; you still do the believing.

The usual approach — paste the PDF into a chat window and ask for a summary — is fast and feels productive. It also gives you a fluent paragraph you cannot evaluate without reading the paper, which is the thing you were trying to avoid.

Why a summary is the wrong first question​

A summary answers "what is in this paper?" in the paper's own framing. That framing is exactly what you are not yet able to judge on a paper outside your field. Worse, a summary has no edges: every sentence sounds equally supported, and the one that is a slight overstatement looks the same as the one lifted straight from the abstract.

The better first question has structure you can check:

What is this paper claiming, and what does each claim rest on?

An answer to that is a short list of claims, each tied to evidence. A list can be checked one item at a time. A paragraph cannot.

The method: the three-pass read, with AI doing the first pass​

S. Keshav's well-known note How to Read a Paper describes three passes: a quick skim for the shape, a careful read of the content, and a deep pass that rebuilds the argument. AI is genuinely good at the first pass and genuinely risky in the third. So split the work that way.

1. Let the AI take the first pass — and make it show its sources​

Ask for the claims and what they rest on. Require verbatim quotes with page numbers. If your tool cannot give you a page you can open, treat its answer as a rumour about the paper.

In MindMax the paper goes into the model whole rather than as a few retrieved fragments, and every quote in the answer is checked against the PDF's text before it is shown. Click one and the reader opens on that page with the sentence highlighted.

A passage sent from the page arrives with its page and section, so the answer can point back to it.

2. Check two claims, not all of them​

Pick the two claims that would change your view of the paper if they were wrong — usually the headline result and the one assumption it depends on. Open both quotes. Read the sentence and the paragraph around it.

This takes about ninety seconds and it is the step that matters. It turns an answer you cannot evaluate into one you have partly verified, and it tells you how much to trust the rest. If both quotes hold up in context, the map is probably good. If one is subtly stronger than the source, you have learned something about the answer, not just about the paper.

3. Read the parts that matter yourself​

Now you know where they are. Read those sections properly and mark them up. This is Keshav's second pass, and it is still yours — the AI has only told you which pages deserve it.

4. Ask about the figure, not the prose​

Papers outside your field rarely lose you in a paragraph. They lose you at Figure 3, or at a column in Table 2 you cannot place. Ask about that region specifically, with the text underneath it, rather than pasting a screenshot into a chat window and hoping the numbers are read correctly. (Why that difference matters is a post of its own: a screenshot is not a region.)

5. Ask what the paper does not establish​

What does this paper explicitly not claim, and what does it never test?

This question surfaces the limitation buried in the discussion section — the one that changes how you would cite the paper. It is also a question that only works if the AI has read the whole paper. Absence does not show up in a few retrieved paragraphs.

6. Leave a trail for next month​

Your highlights and the notes you kept should live with the PDF, not in a chat history. When you come back to cite it, you want to start from where you stopped, not from the abstract.

Where it goes wrong​

  • Dense mathematics. Quotes from heavy notation are the least reliable part of any extraction. Click through to the page rather than trusting the text.
  • Two-column layouts with figures in between. The hardest case for any PDF reader. If a quote reads oddly, the page will settle it.
  • Scanned papers. An old paper with no text layer has to be recognised first; until then an AI answering about it is answering from nothing. How to OCR a scanned PDF covers that step.
  • Judgement. None of this tells you whether the paper is good. It gets you oriented fast, so the time you do spend goes on the parts worth judging.

What is not claimed​

We are not claiming this makes you understand a paper in a set number of minutes, or that the AI's reading is correct. The claim is narrower: when every point arrives with a quote you can open, checking costs a click instead of a re-read — and a check that cheap is one you actually do.

The step-by-step version is understand a paper in about 30 minutes; what the reader does is on the AI PDF reader page, and how quotes are checked is in checked quotes. Download MindMax and open the folder your papers are already in.