How to Research with AI Without Fake Citations

Stop asking a chat model to remember sources, and start giving it sources to read. That single change removes most fabricated citations, because it removes the moment where the model has to invent one. Prompting harder does not fix this. Choosing a differently-shaped tool does.

The failure has a name in court records now. Legal researcher Damien Charlotin maintains a public tracker of court decisions where a party filed AI-hallucinated material and a judge responded to it; by mid-2026 it held over a thousand US decisions and was still growing daily. Penalties have moved past four-figure fines into five-figure sanctions and bar suspensions. Those are lawyers — people whose entire job is checking citations. If it catches them, it will catch you.

Why the citation is fake in the first place

A chat model with no source attached is not looking anything up. It is producing the most plausible next characters, and a citation is one of the most predictable text shapes in existence: author, year, title, journal, volume, pages. The model can generate a perfect-looking one for a paper that was never written, and it will do it with exactly the same confidence it uses for a real one.

That is the important part. There is no tell. The fake citation is not hedged, misspelled, or vague. It is crisp, formatted, and specific, because crisp and formatted is what the model is good at. You cannot spot it by reading carefully. You can only spot it by checking whether the thing exists.

So the fix is structural, not editorial. Move the sources into the tool before you ask the question.

The four postures, and which one you are in

Every AI research tool sits somewhere on one axis: where do the facts come from?

Posture	Where facts come from	Can it invent a citation?	Tools
Closed	The model’s weights	Yes, easily	Claude or any assistant with no sources attached
Retrieval	Live web results it fetched	Real links, but it can still mis-summarize them	Perplexity
Corpus	A curated body of real papers	No — the papers exist by construction	Consensus, Elicit, Semantic Scholar
Grounded	Only the documents you loaded	No — it has nothing else to draw on	Gemini Notebook (formerly NotebookLM)

Closed is not a bad posture. It is the right posture for thinking, drafting, structuring an argument, and pressure-testing your reasoning. It is the wrong posture for “find me sources,” and that mismatch is the whole problem. Most operators use one tool for both jobs and are surprised when the second one fails.

The workflow

1. Split the question into “what do I know” and “what am I arguing.”

Sourcing is an Intelligence-layer job. Arguing is a Creation-layer job. They want different tools and they want them in that order. Doing them in one prompt is what produces a confident essay resting on three papers that don’t exist.

2. Gather sources in a corpus or retrieval tool.

For a research claim — does X actually cause Y — go to Consensus or Elicit. Both read real papers, so the citation problem is answered by construction: they can be wrong about a summary, but the paper is on the shelf. For a current, non-academic question, Perplexity fetches live pages and shows you the links.

3. Load those sources into a grounded tool and ask your real question there.

This is the step people skip. Gemini Notebook only answers from what you gave it — that constraint is not a limitation, it is the entire product. Ask it to synthesize, contradict itself, find the weak point in your argument. Its answers cite back into your own documents, so every claim is one click from the sentence it came from.

4. Verify anything that will leave the building.

Not everything. The claims that carry risk: numbers, dates, quotes, anything with a person’s name on it, anything a reader could act on. Open the source. Confirm the sentence says what the summary says it says.

The rule that survives any tool

Never cite a source you have not opened.

That is the whole discipline, and it applies to a colleague’s memo, a Google result, and an AI summary equally. AI did not create this rule — it just raised the cost of ignoring it, because it made producing an unopened citation take zero seconds instead of ten minutes.

Two habits that make it cheap to keep:

• Keep the link next to the claim while you draft. If a sentence in your draft has no link beside it, it is not a fact yet, it is a placeholder. Strip the links at publish, not before.
• Verify at the point of writing, not the point of publishing. A verification pass at the end is a hundred small tasks you will rush. One check as you write the sentence is one task you will actually do.

What about the model’s own web search?

Retrieval helps and does not finish the job. When an assistant searches the web and cites what it found, the links are real — it fetched them. The remaining failure moves one step down: the summary can drift from what the page actually said, especially when the page is long, hedged, or arguing the opposite of what the snippet implies.

So retrieval converts a hard problem (does this source exist) into an easier one (does this source say this). That is real progress. It is not permission to skip step 4.

The call

Job	Open this
Check whether a research claim holds up	Consensus
Run a real literature review you’ll defend	Elicit
Answer a current question with live links	Perplexity
Synthesize documents you already trust	Gemini Notebook
Think, argue, structure, draft	Claude — with sources pasted in
Find a paper by citation graph	Semantic Scholar

The workflow should choose the tool. If the workflow is “find out what’s true,” a closed chat window is the wrong room to be standing in — no matter how good the model is. Give the question a source to land on, and the fabrication problem mostly stops being yours to manage.

Next: audit your stack to see whether your Intelligence layer has a real seat in it, or read how to get cited by AI search for the other side of this — being the source someone else checks.