Research with AI without inventing sources

Fabricated citations can damage trust. Better prompting may reduce them, but a source-verification workflow is still necessary for claims you intend to rely on.

Without a source-checking step, a generated citation can look plausible whether or not the source exists. Retrieval tools can provide evidence, but the final answer still needs to be checked against what was actually retrieved.

Asking the model not to invent sources can help, but it cannot guarantee compliance. Build verification into the workflow instead of relying on that instruction alone.

Retrieve first, reason second

Retrieve relevant sources first, then ask the model to work from them. This makes the evidence inspectable; it does not remove the possibility of fabricated citations or unsupported conclusions.

Do not ask “what does the research say about X”. Find sources yourself, or use a tool that searches and returns actual documents, then ask the model to work with what is in front of it.

Give it the three papers and ask it to use only those. Then verify that every cited source is among them and that the attributed claim is present. The restriction is an instruction, not a technical barrier to inventing a fourth paper.

Do not use response speed as a source-quality test. Inspect the retrieved documents, citations and supporting passages instead.

Demand verbatim quotes

Ask for quotes you can verify in the actual source.

A short quote and page or section reference make a claim easier to check. Search for the words in the original document and read around them; a quote can be fabricated or real but misleadingly selected.

When a model cannot produce the quote, or produces one that does not appear in the document, you have learned something precise about that particular claim rather than a general unease about the whole output.

Treat every specific as a claim

Check specific numbers, dates, names and titles as well as the overall interpretation. Either a cautious or confident answer can contain a consequential error.

A useful rule: anything specific that you did not supply is a claim until you have checked it. That includes figures inside an otherwise well-sourced summary. Models frequently attribute a real-sounding statistic to a real paper that does not contain it, which is the worst case, because the citation checks out and the number does not.

So verify at the level of the number, not at the level of the source.

Check citations by opening them

This sounds obvious and it is routinely skipped, because a correctly formatted citation looks verified.

Watch for three distinct failures: a nonexistent source, a real source that does not support the claim, and a correct passage used outside its scope. Check all three rather than assuming one is the dominant failure in your workload.

If you take one habit from this post: open the link. Not all of them, if time is short — but every one that is doing real work in your argument.

Ask what it could not find

A question almost nobody asks, and one of the most useful: “What did you look for and fail to find?”

These tools produce complete-seeming answers by default. Gaps in the evidence get smoothed over rather than flagged, because a fluent answer is the default output shape. Asking directly about absence often surfaces that the strongest claim in the summary rests on one source, or that the contrary evidence was not looked for.

If a search turns up little counterevidence, investigate whether the question, search terms or available sources were too narrow. Absence from your search does not establish consensus.

Where AI research genuinely excels

It is worth being clear about this, because over-correcting wastes a real capability.

These tools are excellent at working across material you have already gathered. Finding the tension between two documents. Summarizing twenty items against a consistent set of questions. Noticing that a definition changed between an old policy and a new one. Restructuring a sprawling set of notes into an argument. Drafting the counter-case to something you have written.

Working from a known source set can make review more manageable. You still need to check completeness, interpretation, currency and whether the sources themselves are trustworthy.

The organizational risk is worse than the individual one

An unsupported number can harm an individual decision or spread through an organization. The consequences depend on how it is used.

Consider a hypothetical chain: an unsupported figure enters a slide, the slide is reused and the figure later appears in a board pack without a traceable source. Preserve the source at each reuse so the claim remains checkable.

The defense is a norm rather than a policy: anything that reaches a slide carries a link somebody can open. Not a footnote for form’s sake; a source a colleague can check in thirty seconds. “An AI said so” is not a source, and a team that says this out loud once tends to internalise it quickly.

A note on deep-research modes

Some assistants offer research modes that search sources and assemble cited reports. Treat the report as an aid to investigation, not as independently verified evidence.

Research-oriented tools can retrieve and organize evidence, but a citation is still a claim about a source. Check the important links, exact statements and dates before relying on the report.

Check every claim material to the decision, and expand review according to the consequences. A twenty-page report may contain many important claims; there is no universal three-or-four-sentence checking shortcut.

A workflow that holds up

Gather sources yourself, or with a tool that returns real documents you can open. Put them in front of the model. Ask for analysis with verbatim quotes. Check the quotes for anything load-bearing. Ask what is missing and what the other side would say. Write the piece yourself, keeping the citations you personally verified.

This workflow makes errors more visible and easier to investigate. It does not fail safely by itself: fabricated material can still appear, and an accurate quote can still support the wrong conclusion.

The underlying principle is the same one that governs any number you are going to stand behind — you should be able to say where it came from, and somebody else should be able to follow the same path and land in the same place.

Technical review, October 10, 2026. References: Anthropic: reduce hallucinations; OpenAI: evaluation best practices.

Try this

Inspect three important citations in a recent AI-assisted report: existence, support for the claim, date and context. Record the findings as a spot check, not an estimate of the model’s overall error rate. Expand the review where the consequences require it.