Why ChatGPT Invents Fake Professors and Citations (and What to Use Instead)
ChatGPT confidently invents professors and papers that do not exist. Here is why it happens, how to catch it, and what to use for real research instead.
You ask ChatGPT "which professors work on my research problem?" It gives you five names, their universities, and a representative paper each. It looks perfect. Then you try to email one — and the professor doesn't exist. Or the paper does, but a different person wrote it. Or the paper title is real-sounding and completely made up.
If this has happened to you, you're not doing anything wrong. It's how the tool works.
Why it happens
A large language model like ChatGPT is a next-token predictor. It was trained to continue text in a way that sounds right — not to look anything up. When you ask for professors who work on a topic, it doesn't query a database of researchers. It generates the *kind of text* that would plausibly answer your question: real-sounding names, real-sounding universities, real-sounding paper titles, assembled from statistical patterns.
Most of the time that produces something believable. Sometimes the believable thing is entirely fictional. The model has no built-in sense of which is which, because from its perspective there is no difference — both are just probable sequences of words. Researchers call this hallucination, and for factual lookups (names, citations, numbers) it is not a bug you can prompt your way out of. It is what the tool is.
How to catch a fabricated citation
Before you trust any professor or paper an LLM hands you:
- Search the exact paper title in Google Scholar or the arXiv. If it doesn't return that exact title by that exact author, it's fabricated.
- Check the author is real and at the claimed institution — a university faculty page, not the model's say-so.
- Watch for suspiciously perfect matches. Real search is messy; you get partial fits and adjacent work. A tidy list of five researchers who each perfectly match your niche is a red flag, not a green one.
- Verify the year and venue. Hallucinated citations often pair a real author with a plausible-but-invented paper, or a real paper with the wrong year.
If any check fails, discard the whole answer — one fabrication means the rest is unverified too.
Why "just add web search" doesn't fully fix it
Newer chatbots can browse, which helps. But browsing bolts a search step onto a system whose default behavior is still to generate. It can retrieve a few pages and then confidently summarize them wrong, blend a real source with an invented detail, or fall back to generation when a search comes up thin. Retrieval helps at the margin; it doesn't change the fact that the model's core skill is producing plausible text, not being right about who-wrote-what.
What to use instead: grounding
The alternative is a system built the other way around — retrieval first, generation last. Instead of asking a model to *recall* researchers, you search a real corpus of real papers for genuine overlap, and only then let a language model *read and describe* what was actually found. The model never names a person or cites a paper that isn't in the corpus, because it isn't the thing choosing names — the search is.
That "extract, don't invent" design is exactly what we built Siddhanta around. You give it your paper; it retrieves the nearest real faculty from a corpus of ~45,000 researchers and 1.6 million papers, shows you the specific real paper of theirs that overlaps yours, and if it can't ground a claim, it stays silent instead of guessing. Every claim traces back to a citation you can click.
The rule of thumb for any research question where a name or a number has to be *correct*, not just plausible: use a tool that looks things up, and be suspicious of any tool that's really just predicting text. For finding advisors, reading lists, or citations, "sounds right" is not good enough — and a chatbot is optimized for exactly "sounds right."