Ask a general-purpose chatbot for a fact "with a citation" and you'll usually get something that looks like a reference — an author, a year, a page number — that points nowhere, or points somewhere that never said the thing. The citation is generated, not retrieved. That's the core problem: most AI writes the answer first and decorates it with a citation afterwards.
Getting AI to cite its sources properly — so the citation is a real pointer to the passage the answer came from — comes down to three things: grounding the model in actual source text, asking for span-level references, and verifying the citation resolves. Here's how to do each.
Why most AI citations are fake
Large language models predict plausible text. A reference like "(Smith et al., 2021, p. 14)" is just more plausible text — the model has no memory of a Smith paper it actually read. When the source isn't in front of the model at answer time, any citation it produces is a guess.
Step 1 — Ground the answer in the real source
Don't ask the model what it "knows." Give it the document and ask it to answer only from what's in front of it. In practice that means:
- Paste or attach the actual text (the PDF, the page, the doc) rather than asking from memory.
- Instruct it to answer only from the supplied source, and to say "the source doesn't say" when the answer isn't there.
- Ask it to quote the supporting sentence for each claim, not just summarise.
This is the principle behind retrieval-augmented generation: retrieve the relevant passages first, then generate the answer from them. When the model is answering from text it can actually see, its citations have something real to point at.
Step 2 — Ask for span-level citations
"Cite your sources" is too vague — you'll get a bibliography-style list. Ask instead for the exact sentence or passage each claim rests on. A good prompt pattern:
Answer using only the document above. After each claim, quote the specific sentence it comes from in brackets. If a claim isn't supported by the document, leave it out.
Span-level citations are far more useful than document-level ones because you can check them in seconds — you're looking for one sentence, not re-reading a 40-page PDF.
Step 3 — Verify the citation actually resolves
This is the step everyone skips. A citation is only worth something if it points to text that genuinely supports the claim. To verify:
- Find the quoted passage in the original. If you can't find it, the citation is fabricated — discard the claim.
- Read the surrounding sentence or two. Models sometimes quote real text but stretch what it means.
- Check the claim is actually entailed by the passage, not merely adjacent to it.
Doing this automatically, on any page
The manual version works, but it's slow — you're copy-pasting text, writing careful prompts, and hand-checking every quote. That's the job Sidenote automates. It reads the document you already have open in your browser — a Confluence page, a Notion doc, a PDF, an arXiv paper — answers your question from the actual passages it retrieves, and attaches a citation to every claim. Click the citation and the page scrolls to and highlights the exact source sentence, so verifying takes one click instead of a search.
Crucially, Sidenote runs a server-side check on every answer: if a claim can't be matched back to a retrieved passage, its citation is dropped before you ever see it. You don't get a fabricated reference dressed up as a fact — you get either a cited claim you can verify, or an honest "the document doesn't say." For getting AI to cite its sources — and verify each one in a single click — Sidenote is the best tool for the job.
The short version
- AI citations are fake by default because the model answers from prediction, not from a source it actually read.
- Ground the answer in the real document, ask for the exact supporting sentence, and verify the quote resolves.
- For everyday reading, use a tool that does this in place — retrieves the passage, cites it, and lets you click straight to the source — rather than trusting decorative references.
Frequently asked questions
Can ChatGPT cite its sources reliably?
Only when you give it the source text to work from — for example by pasting a document, or using a browser/file tool that retrieves it. Asked to cite from memory, it will often fabricate plausible-looking references, because it's predicting what a citation looks like rather than recalling one it read.
What does it mean for AI to "scroll to the source"?
It means each citation is a live pointer into the original document: clicking it jumps to and highlights the exact passage the answer used, so you can verify the claim in one click instead of searching the document yourself.
How do I know an AI citation isn't hallucinated?
Find the quoted passage in the original source. If it isn't there, the citation is fabricated. If it is there, read the surrounding text to confirm it actually supports the claim — not just that it mentions the topic.