How to Use AI for Research Without Fake Citations

Why this keeps happening

A general-purpose language model generates text one likely word at a time, based on patterns learned from its training data. When you ask it for a citation, it produces something that looks like a real reference (plausible author names, a real-sounding journal, a year, a page range) because that's the pattern it learned, not because it looked anything up. Most of the time this produces something close to a real paper, or a real paper with a wrong detail. Sometimes it produces something that doesn't exist at all. The model has no internal mechanism telling it the difference, because generating fluent text and reporting a verified fact are two different tasks that happen to look identical on the screen.

This is why the failure is so persistent across tools and model versions: it isn't a bug that gets patched, it's a structural consequence of how the technology works when it isn't given anything specific to check against.

The distinction that actually matters: generation vs. retrieval

The fix isn't "use a better model" or "write a more careful prompt". Those help, but not reliably enough to trust with something you'll submit. The fix is understanding which of two things a tool is doing when it gives you an answer:

• Generation. The tool answers from its general training: everything it read, compressed and blended, with no record of where any specific claim came from. This is what a general-purpose chatbot does by default. It's useful for brainstorming, summarising a concept, or drafting, and unreliable for anything you need to cite.
• Retrieval. The tool answers only from a specific, bounded set of documents (search results it just looked at, or papers you've actually given it) and can point to exactly which passage a claim came from. This is what a citation manager, a search engine, or a properly built research tool does.

Most of the tools people reach for sit somewhere in between, and the marketing rarely tells you which side of the line a given answer landed on. The practical question to ask of any AI research tool, every time: is this answer generated from general knowledge, or retrieved from a specific source I can go and check?

Ask which side of the line the answer came from.

A five-step checklist that works with any tool

Whatever you're using (a general chatbot, a search-grounded assistant, a dedicated research tool), this is the routine that catches fabrication before it reaches your reference list:

• 1. Ask the tool to show its source, not just its claim. If it can't point to a specific document, page, or passage, treat the claim as a lead to chase down, not a fact to cite.
• 2. Search for the reference independently. Title plus first author's surname into Google Scholar, PubMed, or a similar index. If it doesn't surface in seconds, be suspicious rather than assuming your search was clumsy.
• 3. Check the DOI resolves. A fabricated reference often carries a DOI that either doesn't resolve at all or resolves to a completely different paper. It's a fast, near-free check.
• 4. Confirm the claim is actually in the source, not just the source is real. A real, correctly cited paper can still be misquoted or have its finding overstated. Open the passage and read the sentence the claim is supposedly drawn from.
• 5. Never paste a reference list an AI generated directly into your work. Rebuild it from sources you've individually verified, even if that means retyping details you already had right.

None of this is exotic. It's the same scepticism you'd apply to an unfamiliar source in any other context. It just has to become a fixed habit, because the fluency of the output makes it easy to skip.

What a retrieval-grounded tool actually buys you

A tool built around retrieval doesn't eliminate the need to check (no tool does), but it changes what you're checking. Instead of verifying that a reference exists at all, you're confirming that a real passage, in a document you already trust enough to have uploaded, says what the answer claims it says. That's a faster check, and it's the one this kind of tool is built to make easy: a citation that opens directly to the highlighted passage in the source PDF, rather than a reference you have to go and hunt down from scratch.

This is the structural idea behind MyOwnArc: it answers only from papers already in your library, and every claim carries a citation that opens the exact passage in the source PDF, so checking a claim is a two-click habit instead of a separate research task. It doesn't replace the checklist above; it's what makes step 4 (the one that actually catches misquoting, not just fabrication) take a few seconds instead of a few minutes.