How to Research With AI Without Inventing Sources

Ask a model for sources on a niche topic and you may get three references back: plausible authors, a real-sounding journal, a sensible year, a title that fits the argument perfectly.

None of them exist.

This is not a bug being fixed next month. It follows directly from how the thing works — it predicts what a citation should look like, and it has seen a great many citations. It has learned the shape of a reference, not the contents of a library.

The structure that fixes most of it

Do not ask "what does the research say about X". Ask for the answer split into three parts, kept separate:

Part 1 — what is well established. Only things it is confident are widely accepted. No invented sources; if it cannot name one, it should say so.

Part 2 — what is contested. The competing positions, and who holds them.

Part 3 — what I should verify myself. Exact search terms or document names, and what a good answer would look like.

The separation is the whole trick. A blended answer hides which parts are solid. Three labelled parts make the shakiness visible, and a short Part 1 is itself a useful finding — it tells you the ground is less settled than you assumed.

Treat every citation as a search term

Never cite something the model gave you without finding it yourself. Not "check it looks right" — actually locate it.

The useful reframe: a citation from an AI is a search term, not a source. It is often a good search term. It points you at roughly the right literature. But you are the one who has to go and find the real thing, and sometimes what you find says something different.

Prefer tools that can search

If the tool has web access, use it and ask for links. Retrieval-augmented answers — where the model fetches material and then writes from it — are meaningfully more reliable than answers from memory alone, because the claims are anchored to something you can open.

Even then, open the links. A model can misread a page it has genuinely retrieved.

Where AI is actually excellent at research

The scepticism above is about facts. There is a lot of research work where it is genuinely strong:

• Summarising material you supply. Paste the report; ask for the argument.
• Finding the shape of a field — the main positions, the vocabulary, what to search for.
• Generating angles. Twenty questions you had not thought to ask.
• Explaining something dense in plain language so you can decide whether to read it properly.

Notice that these are all cases where you supply the ground truth or verify it immediately. That is the pattern worth internalising: use it to process information you have, and be careful when asking it to supply information you do not.