Can You Trust AI Citations? How to Check a Reference
· 6 min read
Not without checking. An assistant that is not connected to a real search, or to documents you gave it, can write references that look right, with author names, a title, a journal, a year and even a link, yet do not exist or do not say what it claims. Even a real source can be cited for a claim it never makes.
Why do AI assistants make up citations?
A language model produces fluent text that fits the context, and a reference has a familiar shape: a couple of surnames, a title that sounds like the field, a journal name, a year, page numbers. It can write that shape without looking anything up, filling it with plausible parts when nothing real is in front of it.
A made-up reference arrives in the same calm tone as a real one, so a convincing look tells you nothing. Why AI makes things up explains the wider problem, often called hallucination.
In what ways can a citation be wrong?
There are four, and each needs a different check.
- The source does not exist. The title, authors and link are invented, or stitched together from real parts, such as real authors with a title they never wrote.
- The source exists, but the details are wrong. The title is close, yet the year, the journal, an author or the link belongs to a different item.
- The source exists but does not support the claim. It is on the topic, yet says something weaker or narrower: a study of children cited for a claim about adults, or "may be linked to" turned into "causes".
- The source is real and relevant but out of date. It may have been corrected, replaced by newer findings or withdrawn, and the assistant may not mention it.
The second and third are the hardest to spot: a title search finds something, and it is easy to stop there.
How do I check an AI citation step by step?
Work through these in order, and stop trusting the citation at the first clear failure.
- Search the exact title in quotes in an ordinary web search, not in the assistant. A real item usually appears under that title. A missing match is a warning but not proof, since some items are not online.
- Confirm author, year and venue. Names, year and journal should match. Real authors with the wrong title suggest an assembled reference. Follow any link or DOI (a paper's permanent ID) and check that it opens that item; a working link can lead to a different one.
- Open the source and find the sentence. Search the page (Ctrl+F or Cmd+F) for the claim's key words and check that the strength of the wording and the group studied match. If you can only reach an abstract, say your check covered the abstract.
- Check the date and any notices. Look for the words correction, retraction or withdrawn near the title, and for newer work that replaces it.
- Look at who published it. A journal, a government agency, a company selling a product and an anonymous blog carry different weight. How to fact-check an AI answer covers preferring the original source.
What does checking a made-up reference look like?
Suppose you ask an assistant whether short breaks help study focus. It replies that regular rest intervals improve sustained attention in adult learners, and gives this reference: Marlowe and Eastgate (2019), Interval Rest and Sustained Attention in Adult Learners, Journal of Applied Learning Behavior, 14(3), 211-229. Every part of that reference is invented for this example.
Say a search for the title in quotes finds nothing, and without quotes only general articles on breaks. Checking the parts goes no better: the journal name matches nothing, and one surname belongs to a researcher in an unrelated field. The reference fails, so the claim has no support yet. It may still be true, but you have not learned that. Search for the claim itself in plain words and look for sources you can open.
Now suppose the title did match a real paper, but under a different year, and its abstract describes schoolchildren, not adults. A title search alone would not catch that.
Are citations from an assistant with search more reliable?
In one respect, often yes. When an assistant has a working search tool, its links usually point to pages it retrieved, so the source probably exists.
It does not fix the third failure, a real source that does not support the claim. The assistant can still misread a page, attach a real link to a claim the page does not make, or rely on an unreliable page. A link beside an answer shows where it says it looked, not that the page agrees. Some assistants answer from memory yet still write what looks like a source list, so if you cannot tell whether it searched, assume it did not.
What should I do when I cannot find the source?
Treat the claim as unsupported, and say so. Do not let it into an essay, a report or a message as if it had been checked.
- Do not ask the assistant to confirm it. "Is this reference real?" can bring a yes, or a second invented reference. Only a source you opened counts.
- Look for the claim, not the reference. Search the claim in plain words for official or original sources you can open, and cite what you actually found.
- Mark what remains. If nothing turns up, write "I could not verify this" in your notes, or drop the claim.
- Raise the bar with the stakes. For health, money, law or safety, rely only on an official source or a qualified professional.
A real source you cannot open, such as a paywalled paper, is different: you can confirm it exists, not what it says. Say exactly that.
How do I ask an assistant for citations I can check?
- Give it the document and ask for the exact sentence. "Quote the sentence from the text I pasted that supports each claim, and say where it appears." Then search your file for those exact words: a quote can be altered or invented even with the text in front of it.
- Allow "I do not know." "If the text does not support a claim, say so instead of guessing."
- Separate your document from general knowledge. "Label each statement as from my document or from general knowledge, and give a reference only for the first kind." That may cut down on invented sources, but the labels can be wrong too.
When a detail matters, read the source itself; AI summary vs reading the source shows where that line falls. For a whole project, a method for researching with AI keeps the evidence in view, and how to read a research paper with AI goes through a single paper.
Where NAMELESS stands
NAMELESS is an AI model still in training. It is not yet available and has no launch date, so you cannot try any of the checks above on it today. Our goal is a model on our own architecture, with unlimited context and no rereading of the conversation, extremely fast, with stronger logic and reasoning. None of that removes the need to open the source and find the sentence yourself. If you would like to follow along, join the waitlist.