How to read research papers with AI
Published · by the team that makes Plum
An AI assistant can get you through a paper much faster, but only if you keep checking it against the paper. This method lets the AI speed up the reading without doing the reading for you.
The one rule: ask for evidence, not summaries
A summary you can’t trace is a guess. Ask questions whose answers point to a passage, figure, or equation, then look at it. If an answer can’t point anywhere, treat it as unverified. Everything below follows from this.
Three passes, with an assistant
S. Keshav’s widely used guide “How to Read a Paper” splits reading into three passes: a quick survey, a careful read, and a deep one. Each pass has a different job for the AI.
Pass 1: decide whether to read it
Read the title, abstract, introduction, section headings, and conclusion yourself; it takes a few minutes and gives you the shape of the paper. Then use the assistant to fill in what you skimmed past:
- “What problem does this paper say it solves? Show me where.”
- “List the contributions the authors claim, quoted.”
- “What data, methods, and baselines does it use?”
Check: open each passage it cites. If the claimed contribution isn’t in the text, the answer is wrong.
Pass 2: understand it
Now read the paper through, using the assistant where you get stuck rather than in place of reading:
- “Explain this equation term by term, and where each term is defined.”
- “What does Figure 3 show, and which experiment produced it?”
- “Where is this term first defined?”
- “What does reference 12 claim, and how does this paper use it?”
Check: compare each explanation with the passage. Where the paraphrase and the paper disagree, the paper wins.
Pass 3: judge it yourself
Judging a paper is the part to keep. Write down what you think it shows and how well, then use the assistant as a search tool against your own view:
- “Where do the authors discuss limitations?”
- “Find every place the main result is qualified or hedged.”
- “Find passages that contradict this summary: …”
Check: the assistant finds passages; you decide what they mean.
Where AI misleads on papers
- Confident summaries of things the paper doesn’t say, or says only for one setting.
- Blending cited work with the paper’s own claims, so a related paper’s result is credited to this one.
- Dropping hedges like “may”, “in our experiments”, or “preliminary”, which turns a tentative result into a firm one.
- Misreading figures, tables, and equations, which are harder to extract from a PDF than prose.
- Reading part of a long paper and answering as if it read all of it.
Doing this in Plum
We built Plum around this way of reading. Ask a question and the assistant turns to the page and highlights the exact lines it used, so every answer comes with its evidence. Search a paper by meaning, exact text, or LaTeX; hover a citation to preview the reference; copy any equation as LaTeX. Papers are processed on your Mac, and you choose the AI: the built-in one, your own key, or Claude over MCP.
More on Plum for researchers, or compare the best AI PDF readers for Mac.
Questions
Can AI read a research paper for me?
It can summarize one, but a summary you haven’t checked can be confidently wrong. Use AI to find and explain passages faster, and keep the judging for yourself.
What is the best AI for reading research papers?
One that shows its evidence. Whatever model you use, prefer a tool that points to the exact passage behind each answer. See the best AI PDF readers for options.
Is it safe to give unpublished papers to an AI tool?
It depends on the tool. Most chat-with-PDF tools upload the whole document. Plum processes PDFs on your Mac and sends only the passages the assistant reads to the AI you choose.