Skip to content
Theoria home
Log in

For students, teachers and curious readers

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

AI can turn a dense paper into plain language in seconds. That is useful, and it is risky: a fluent summary can hide what a study actually measured. This guide shows a way of reading with AI that keeps you in control of the claims.

Updated on 20 October 2026

The question

How can I use AI to understand a scientific paper without simply believing the summary?

Step by step

  1. Step 1

    Start with the original

    Read the title, the abstract and the first figure yourself before you ask any AI. Write down in one sentence what you think the study asks. Experienced readers work in passes like this: first an overview, then the details (S. Keshav describes a three-pass method).

    In Theoria: Every paper opens with a simplified summary, and the original abstract is one click away. Read both and compare them with your sentence.

  2. Step 2

    Identify the type of study

    Is it an experiment, an observational study or a review? Is it a preprint or an article in a peer-reviewed journal? The design limits what a paper can show: an observational study can find associations but cannot prove causes. Also check whether there is a newer version or a published correction.

    In Theoria: Ask Albert what kind of study it is and what this design can show. Then check the answer against the methods section.

  3. Step 3

    Find the numbers behind the main claim

    Who or what was studied, how many, for how long and compared with what? Note the sample size, the comparison groups and the main measurement. These few facts decide how far a result can be generalised.

    In Theoria: Deep Insights explains a paper at a depth you choose from 1 to 10: a low depth gives an overview, a higher depth explains methods and background.

  4. Step 4

    Ask targeted questions

    Instead of "summarise this", ask: Which result supports the main claim? How was it measured? What do the authors say about limitations? Ask where in the paper the answer comes from and look the passage up yourself.

    In Theoria: Albert answers in the context of the open paper. If an answer has no clear basis in the text, treat it as a lead, not as a fact.

  5. Step 5

    Look for limitations and other studies

    Read the discussion and limitations yourself. Then look for later studies, replications or criticism on the same question. One paper is rarely the whole answer.

    In Theoria: AI Peer Review helps you list the strengths and weaknesses of a paper. Deep Answer draws on several studies for one research question and links the sources.

  6. Step 6

    Write your own three-sentence summary

    What was studied, what was found, and what the study does not show. Compare it with your sentence from step 1 and with the AI summary. Where they differ, go back to the paper.

Worked example

A preprint about essay writing with ChatGPT

In June 2025 a research team with Nataliya Kosmyna as first author posted a study on arXiv titled "Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task". The title sounds like a verdict. Here is what the steps reveal.

Type of study
A preprint: publicly available, but the arXiv entry lists no journal publication (checked on 20 October 2026). A revised version 2 appeared on 31 December 2025.
Who and how many
54 participants took part in sessions 1 to 3, divided into three groups: with an LLM, with a search engine, or without tools. 18 of them completed a fourth session in which the conditions were switched.
What was measured
Brain connectivity with EEG during writing, the essays themselves (language analysis and scores from teachers and an AI judge) and how strongly participants felt the essays were their own.
What the abstract reports
The LLM group showed the weakest connectivity, reported the lowest sense of ownership of their essays and had trouble quoting their own work.
What it does not show on its own
Whether the effects hold for other groups, other tasks or other ways of using AI. The results of the switch session rest on 18 people.

The study raises a serious question and is worth reading in full. A sentence like "ChatGPT damages the brain" would go far beyond what one writing task with 54 participants can show.

Where AI reaches its limits

  • AI summaries can be fluent and wrong. They can merge details, drop a limitation or state a result more strongly than the authors do.
  • AI cannot judge the data better than the paper describes it. If the methods are unclear, the AI is guessing too.
  • When you quote or cite, use the original paper, not a summary.
  • Theoria's explanations are generated by AI and can contain mistakes. Check important claims against the linked source.

Try it with a paper that interests you

Search for a topic, open a paper and compare the simplified summary with the original abstract. Albert, Deep Insights and AI Peer Review need a free account; some tools have monthly limits on the free plan.

Start for freeSources for a research paper

Sources

  1. Kosmyna, N., Hauptmann, E., Yuan, Y. T., Situ, J., Liao, X.-H., Beresnitzky, A. V., Braunstein, I., & Maes, P. (2025). Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task. arXiv:2506.08872. https://arxiv.org/abs/2506.08872Preprint, version 2 from 31 December 2025. Basis of the worked example.
  2. Keshav, S. (2007). How to read a paper. ACM SIGCOMM Computer Communication Review, 37(3), 83-84. https://doi.org/10.1145/1273445.1273458A short, widely cited three-pass method for reading research papers.

Publication details checked against Crossref and arXiv on 20 October 2026.

Other guides

All guides