Searching for papers with ChatGPT —without being fooled by papers that don't exist

Conversational AIs like ChatGPT don't answer by querying a bibliographic database (some have search features, but their basic mechanism is different). Because they work by generating plausible text, they sometimes assemble plausible-looking bibliographic details with no connection to anything real.

The author is a real researcher, the journal is a real journal — but no paper with that exact combination exists. It's a classic, and genuinely troublesome, pattern.

Let's be clear first about the use to avoid.

Asking it to "make me a list of papers" and dropping that list into your references without checking a single title. Using ChatGPT's bibliographic output without verification is flatly wrong. Citing literature that doesn't exist is a fatal error in a report and in a paper alike.

So what is it for? First: preparing your search.

"I want to find papers on this clinical question. Suggest English search terms, including synonyms." For this purpose, it doesn't matter whether the output is factually grounded — because every suggestion is destined to be tested against a real database and kept or discarded on the results.

Second: organizing your question. For example, "structure this question as PICO," or "what should I pay attention to when reading this paper?"

As a thinking partner before the search, conversational AI is genuinely excellent. Not "searching," but "preparing to search" — that, I'd argue, is the right relationship between ChatGPT and paper searching.

Whenever a specific paper title comes up in a ChatGPT conversation, verify it exists before using it.

Search the title in PubMed or Google Scholar. If nothing comes up, try again with the author's name plus keywords. If it still can't be found, treat the paper as nonexistent. And even when it does exist, confirm the bibliographic details — authors, journal, year — as they appear in the database.

Once a paper checks out, get the PDF and move on to actually reading it.

Upload it to Paperfy and the AI summary, figures, tables, and original text come together on one page per paper. Read the original text through the "reading angles" you developed in conversation with ChatGPT, and edit the summary by hand to preserve what you learned. That's how the fruits of the conversation enter your library as verified knowledge.

Put together, it looks like this: ChatGPT = pre-search advisor (search terms, PICO structuring, reading angles). PubMed and Google Scholar = existence checks and the real search. Paperfy = the post-search library (reading, verifying, keeping).

Don't let the three roles blur. That alone makes paper searching in the AI era both safer and faster.