Why your AI assistant keeps making up citations
If you've asked ChatGPT or Claude for a source and gotten back something that looked completely legitimate — real author names, a plausible journal, a DOI-shaped string — and then couldn't find it anywhere, you're not imagining it. This isn't a rare glitch. Depending on the study you look at, something like one in five citations from a general-purpose AI chatbot points to a paper that doesn't exist.1
Here's the part that's easy to miss: the fake ones don't look fake. A hallucinated citation from a language model is built the same way a real one would be — plausible author, plausible year, a journal that actually exists, sometimes even a DOI format that's technically valid, just pointing at nothing. The model isn't lying to you on purpose. It's predicting what a citation should look like based on patterns, not pulling an actual record from anywhere. There's no database check happening in the background unless you specifically use a tool that does one.
And it's not just an "AI chatbot" problem anymore. An analysis of PubMed-indexed papers found fabricated references in roughly 1 in 277 papers published in early 20262 — meaning this is already leaking into published, peer-reviewed work, not just draft lit reviews.
The fix isn't "don't use AI to help with research." It's knowing where the line is between AI suggesting a direction and AI inventing a fact. A language model is genuinely useful for summarizing a paper you already have open, explaining a concept, or drafting a paragraph around a citation you already verified. What it shouldn't be doing is the last step — confirming that a source is real. That part needs to touch an actual database.
That's the one thing Scout does and nothing else: you type a topic, it queries 10 live scholarly databases (Semantic Scholar, OpenAlex, CrossRef, arXiv, and others) at the same time, and only shows you what those databases actually returned. If a paper isn't in any of them, it isn't in your results — there's nothing for the model to "smooth over" or guess at. (What each of those databases covers, and where each has gaps, is its own question — that's a separate guide.)
If you already have a bibliography and just want to know whether anything in it is fabricated, duplicated, missing (cited in your text but absent from the .bib), or orphaned (the reverse — sitting in the .bib but never cited), that's a five-minute job for BibCheck rather than something to catch by manually re-reading fifty entries two days before a deadline.