Summarizing papers with ChatGPT:ask better questions than "summarize this"

Ask an AI to simply "summarize" and it tends to return a generic condensation—something close to a paraphrase of the abstract.

It's not a bad answer, but it flattens the structure of the research: what the authors set out to test, how, what they found, and where the weaknesses are. The shallowness isn't in the AI's capability—it's in the resolution of the request.

My recommendation: split your request for a single paper into smaller questions. For example: "What are the aims and hypotheses of this study?", "Who were the participants and what were the methods?", "What are the main results, with the numbers?", "What limitations do the authors acknowledge?", and "What should I highlight if I present this?"

Smaller questions improve both the accuracy of the answers and your ability to verify them. Once you think of a summary not as one monolithic block but as something you receive in parts, general-purpose AI becomes a genuinely powerful partner.

However well you phrase the prompt, the output can still round numbers incorrectly, misread a comparison, or state conclusions more forcefully than the original does.

For any number, any clinical or research conclusion, and anything going into slides or an abstract, always go back to the original PDF. AI summaries are the entry point; the original paper is the evidence. No prompt technique changes that.

The other problem is storage. Even a well-crafted summary becomes nearly impossible to find in your chat history a few weeks later.

Worse, because the summary and the PDF live in different places, the moment you decide to "check the original," you first have to hunt for the file. A storage setup that makes verifying the original this hard is self-defeating for academic reading.

There's also a family of "pasting problems" that degrade summary quality: two-column layouts getting scrambled, references bleeding into the text, or the second half getting cut off when you copy from a PDF.

We cover this in detail in our article on summarizing PDFs with AI—if pasting is your preferred workflow, it's worth a read.

If you read papers regularly, it's worth adopting a setup where summaries are stored together with the papers themselves.

In Paperfy, uploading a PDF puts the AI summary, figures, tables, and original text on a single paper page. The summary is hand-editable, so corrections you make after checking the original are saved in your library instead of vanishing into a chat thread. Radio scripts are editable too, and you can regenerate the audio from your revised script.

There's no need to abandon general-purpose AI. For in-the-moment interactions—thinking through questions as you read, getting a statistical method explained, checking phrasing in another language—chat AI shines.

Keep the summaries worth keeping in a library; ask the throwaway questions in chat. That division of labor may be the most practical way to read papers in the age of generative AI.