How to Verify AI Research and Catch Made-Up Citations

AI research reads well even when it is wrong, and people trust it. In a 2025 study by 22 public service broadcasters in 18 countries, 45% of AI assistants' answers about the news had at least one significant issue, and 31% had significant sourcing issues (EBU/BBC report). A confident answer built on made-up sources is the same risk in research form.

To verify AI research, check three things for every claim that matters: the source exists, the source says what the answer says, and the claim is current. You can do that by hand for a few claims. At research volume, the checks have to be built into how the research is produced.

Are you interested in more? Book time with us!

Why AI research invents citations

Language models produce text that looks right, and to a model a citation is more text. NIST calls the failure "confabulation": the production of confidently stated but erroneous or false content, known colloquially as hallucinations.

Three conditions make it worse:

• Narrow fields. The fewer real sources exist on a topic, the more the model fills the gap.
• Time limits. Questions like "what changed since March" ask for facts the model may never have seen.
• No retrieval step. An answer written from memory has no document behind it to check.

FifthRow sees this in its own research runs. On questions with timeframes in a narrow field, citation hallucination can be up to 50% without Answers, the agent that strips made-up citations.

The three checks every cited claim needs

A claim is verified when you have confirmed the source exists, says what you quote, and still holds. Skip any one of the three and the claim is unverified, however good it reads.

1. The source exists

Open the link. Check the title, author, publisher and date against what the answer gives. A real journal name attached to an article that does not exist is a common pattern.

2. The source says what the answer says

Find the exact sentence or figure on the page. Paraphrases drift: "up to 30%" becomes "30%", a forecast becomes a result, a US figure becomes a global one. Copy the exact words into your notes next to the claim.

3. The claim is current

Check the date of the source and whether a newer edition replaced it. A 2023 survey cited as "the latest figure" is a quiet error that no link check catches.

A verification checklist for AI research

Use this before any AI-assisted research leaves your team. It works whether the research came from a general AI assistant, a research agent or an analyst.

The last row matters most. A research answer that never says "no source found" is hiding its gaps.

What to build in when research runs at volume

Manual checking works for one memo. It breaks when a team runs hundreds of research questions a month, because the checking takes longer than the research. At that point, verification has to be part of the system that produces the answers.

Four controls do most of the work:

1. Retrieval first. Every answer starts from documents found for that question, not from model memory.
2. Citation checking. Each cited source is fetched and matched against the claim before the answer is returned. Citations that fail are removed.
3. A facts layer. Verified facts are stored once with their source, so the next question reuses them instead of finding them again.
4. A reviewer agent. A separate agent reads each output against a quality bar: what was left out, where each claim came from, and whether it is assumed or triangulated.

This is how the FifthRow platform is built. Answers searches multiple engines and strips made-up citations. Fact checking turns sources into reusable facts. Quality control is a judge agent that reads every output before it leaves.

What verification cannot do

Verification proves a claim is sourced. It does not prove the source is right, and it does not make the judgement for you.

Keep people in charge of:

• Choosing which sources count. A vendor survey and a statistics office are both "sources". Your team decides which ones carry a decision.
• Weighing conflicting evidence. When two good sources disagree, a checker can flag it. A person has to decide.
• Signing off. The person whose name is on the recommendation reads the high-impact claims against the source.

A worked example: research rated above the existing tools

A $1B+ global premium appliance manufacturer runs a small emerging technologies lab that answers research questions for business units across the group. Before, every request landed on one small team, and manual searches of a startup database returned little.

FifthRow set up a research layer the lab operates itself: a question goes in, a source-traced analysis comes out, and the method is saved for the next question. The lab has answered 7,641 source-verified research questions on it. In the group's comparison of AI tools, it was rated above the general AI tools already in use.

We deliver the outcomes, or you run the work on FifthRow. If you are weighing whether to build this kind of system yourself, see our piece on the build vs buy decision for multi-agent platforms.

Are you interested in more? Book time with us!

Frequently asked questions

How do you verify AI research?

Check every claim that matters three ways: the source exists, it contains the exact figure or sentence quoted, and it is current. Prefer primary sources such as regulators, statistics offices and peer-reviewed papers. At volume, build these checks into the research system instead of doing them by hand.

Why does AI make up citations?

A language model writes text that fits the pattern of an answer, and a citation is part of that pattern. Without a retrieval step and a check against the real document, nothing stops it from writing a plausible reference that does not exist. NIST calls this confabulation.

Can AI check its own citations?

A separate checking step can: fetch each cited source and confirm it contains the claim. A model asked to "double-check" its own answer from memory is far less reliable. Keep a person on the final read for high-impact claims.

What is the difference between a verified and a triangulated claim?

A verified claim has a source that exists and says what you quote. A triangulated claim is supported by two or more independent sources. Decisions with real money or risk attached should rest on triangulated claims.

The takeaway

AI research is only as good as its checks. Confirm that each source exists, says what you quote and is current, and build those checks into the system once research runs at volume. People still choose the sources and sign off.