An AI citation is fake in two cases: when the reference does not exist, and when a real reference is tied to a claim it does not support. So check four things, one at a time. Does the source exist, and do its author, title and date match? Does the source say the thing, and is it the right authority for the claim? Most people only check the first, and in 2026 that is no longer enough.
The reason is a shift in what fabricated citations look like. The main 2026 failure mode is not a broken link, but a real, live identifier sitting next to a title that does not match the paper it points to. Most writers rely on one check: paste the DOI, see if it resolves. That check passes this case. This guide walks through the four checks, then the workflow that catches all of them, then the recent cases that show what skipping it now costs.
How do you check if an AI citation is fake?
Run every AI-generated citation through four checks, in order, and the citation has to pass all four. A failure at any step means one thing: do not publish the claim as cited.
1. Does the source exist?
Search the title and author in the canonical database for the field. That is Google Scholar or the publisher site for academic work, PubMed for biomedical, the court reporter or PACER for legal, and the outlet's archive for journalism. Zero results means fabricated, and so does a title-author pair that does not co-occur. This catches the easy half of the problem.
2. Do the author, title, and date match?
Open the actual record. Confirm the author list and the exact title, the year, and the volume and page, or the docket number. A 2026 fabrication often keeps a real identifier and changes the title or authors around it. Say a field disagrees with the live record: treat the citation as fake, even if something at that identifier exists.
3. Does the source actually say it?
Pull the source and find the passage that supports the claim. Search the quoted language. A real paper may not contain the claimed finding, and that is still a failed citation. This is the check almost no one does, and it is also the one that fails most on confident-sounding AI drafts.
4. Is it the right authority for the claim?
A blog post cited for a clinical fact. A withdrawn paper. A press release standing in for the study. A lower court overruled on appeal. All of them are real sources used wrongly: the citation resolves and the metadata matches, and it still does not carry the weight the claim needs.
Why does a DOI that resolves no longer prove a citation is real?
Through 2025, the fast check for a fake reference was simple. Click the DOI or the link: if it resolved to a real page, the citation was probably fine, and if it 404'd, it was fabricated. The most common 2026 pattern defeats that check. The pattern is a real, live identifier bound to a title that does not match the paper it points to. The DOI works and the paper at the DOI is real, but it is just not the paper being cited. So click through and see a valid article, and that tells you nothing: the article you land on is not the one in the reference list.
The fix is a title-to-identifier cross-check, not an identifier-resolves check. Resolve the identifier first, then confirm the title, authors, and year on the landing page: they must match the reference character for character. The mismatch is the tell. Some screening only confirms that identifiers resolve, and it will pass fabricated citations. Other screening flags on surface patterns, and it will wrongly accuse real ones. The failure is documented in both directions, and covered in why AI citations keep showing up wrong.
What do fake AI citations actually look like?
There are four distinct shapes, and they need different checks, which is why a single test misses most of them.
Fully fabricated
No such paper, case, or article: a plausible title, real-sounding authors, standard volume and page numbers, and nothing behind it. Caught by check 1.
Real source, wrong metadata
A real identifier with the title or authors altered around it. It resolves cleanly, and it fails a character-level metadata match. Caught by check 2.
Real source, claim not in it
The paper exists and the metadata is correct, but it does not contain the finding the draft puts on it. That is the hardest to catch, and the most common in long AI-assisted drafts. Caught by check 3.
Real source, wrong weight
It exists, it matches, and it may even support a version of the claim. But it is a preprint, a retracted paper, a secondary summary, or an overruled ruling. Caught by check 4.
The step-by-step workflow to verify AI citations
Run this on every AI-assisted draft before it leaves your desk. It is the manual version, and the end of this guide covers how to compress it.
Step 1: Extract every citation and the claim it supports
List each reference next to the exact sentence it backs. You are not checking citations in the abstract; you are checking that this source supports this claim. Citations with no claim next to them are filler, so cut them.
Step 2: Triage by stakes
Mark the claims that would embarrass you if wrong, or cost a client, or move a decision. Statistics, legal holdings and medical guidance go first, and so do attributed quotes and superlatives. A wrong adjective is survivable. A wrong number under your name is not.
Step 3: Resolve and match, not just resolve
For each citation, open the canonical record and confirm it exists. Then match author, title, year, and locator against the live page, character for character. Treat any mismatch as a fabrication, not a typo.
Step 4: Read the source for the claim
Find the passage that supports the claim. You may not find it in under a minute, and then the burden is on the citation, not on you. Quote-search the source text, and do not trust a summary.
Step 5: Publish, revise, or remove
Each claim gets one of three outcomes: keep it, because it checked out, or weaken it to what the source supports, or remove it. There is no fourth option called probably fine.
Notice where the time goes. Steps 3 and 4 grow with the number of claims, and they are exactly the steps a single AI cannot do for itself. The model that produced the citation has no way to know it is not real, and that is the gap TrueStandard was built for. Paste the draft, and four to five models from different vendors check every claim in parallel, in about 60 seconds. Every citation they disagree on is surfaced for you to read yourself.
We tested that last point directly. Four AI models graded the same answer twice, once with two working DOIs, once with two dead ones, and none of them could tell which was which. So step 3 has to be a lookup, not a question you put to the model.
What are the warning signs of a hallucinated citation?
Before the full workflow, scan for these. They do not prove a citation is fake, but they raise the odds enough to check it first.
- ▸ A title that is too on-the-nose for the exact claim, as if written to order.
- ▸ A precise statistic with a single citation and no methodology anywhere near it.
- ▸ A bare DOI or URL with no author or venue, or an identifier whose landing-page title you have not read.
- ▸ An attributed quotation you cannot find by searching the source itself.
- ▸ A reference list that is oddly consistent in format, where real lists put together by a human are messier.
- ▸ A confident, specific claim sourced to an obscure or hard-to-access work, one that conveniently cannot be checked in a hurry.
What happens if you publish unverified AI citations?
In the 30 days before this guide, the cost stopped being a guess, and it landed across academia, law, and journalism.
arXiv moved to ban it
arXiv announced a one-year submission ban that applies to authors who submit papers with hallucinated citations. The research community framed it this way: writing with AI is allowed, and the author is fully accountable for catching errors of this kind. (Hacker News discussion, May 15 2026)
The scholarly record is being audited retroactively
An audit covered roughly 2.5 million biomedical papers and found on the order of 3,000 with fabricated references, part of a documented multi-fold rise since 2023. Screening old work is now standard, so an unverified citation is no longer a private mistake: it is one someone can find. The full evidence base is in why AI citations keep showing up wrong.
Sanctions, not corrections
A Nebraska brief reportedly contained 57 fabricated citations out of 63, and elite firms have said sorry to federal judges for AI-introduced errors. Courts issued multiple sanctions in a single month, and the rule that came out of it is California's verify-every-AI-output rule.
Bylines and jobs
A senior reporter was dismissed after fabricated quotes reached print, and a newsroom executive was suspended over the same failure. Under a byline, an unverified citation is a career risk, not an editing note.
One caution matters for your own credibility: automated screening produces false positives too. Researchers have had real, well-formatted, DOI-backed citations wrongly flagged as fabricated by LLM-based reviewers, and they then spent weeks proving the references were genuine. So the goal is proof you can show, not a single tool's verdict. And the problem underneath is structural rather than fixable by a better single model.
Do citation-checking tools actually catch fake AI references?
A class of reference auditors turned up in 2026: paste a reference list, get a real-or-hallucinated verdict per entry. They help with check 1 and parts of check 2. They have three limits, and you should know them before you rely on one.
They rarely verify claim support
Most tools confirm a reference exists and that the metadata matches. Almost none read the source, so almost none confirm it supports your sentence. Check 3 stays manual.
A single-model checker inherits single-model blind spots
Ask one model whether a citation is real and you are asking the same kind of system that made it up, whose errors line up with its own. This is the closed loop explained in multi-agent versus multi-model.
They flag real work as fake
Surface-pattern detectors wrongly accuse genuine citations, most often in niche subfields. Some tools cannot show you why they flagged something, and those can damage a real reference list.
Here is the practical reading. A reference auditor is a triage filter, not a sign-off. It tells you where to look first, and it does not replace reading the source for the claims that matter. The free AI citation checker runs checks 1 and 2 as lookups, and check 3 as a reading step that has to quote the passage. So you can see why each reference passed or failed. The same check reads an AI answer against policy pages you choose. See how to verify AI answers against your sources.
What should you check before you publish, by content type?
The four checks are constant. Format changes one thing: which citations carry the most risk, and so go first.
Blog posts and SEO content
Statistics and the named studies behind them come first, because they get screenshotted and they are what rivals fact-check. Check every stat to its primary source, not to the second-hand article that quoted it. A cited round number with no method behind it is a flag, not a fact.
Newsletters
Attributed quotes come first, and so does any claim a subscriber could reply-all to correct. A wrong reference here costs you a public correction email to your whole list, so sort toward anything a reader would know and challenge.
B2B case studies and whitepapers
Customer numbers, market-size figures, and claims about rivals: these move deals. In a review they invite one question: where did this number come from? Keep the log of what you checked, and sign-off does not become a bottleneck.
Technical explainers
Version-specific facts, API behavior, and benchmark figures: these age fast, and here almost-right is wrong. Confirm the claim against current primary docs, not against a model's memory of them.
How do you make citation checking fast enough to actually do?
The workflow above has an honest problem: it is the thing AI was supposed to save you from. Before AI, checking each citation by hand took the better part of an hour per piece. That is the hidden reason AI drafting felt fast: the work moved downstream, and it did not go away. This is the bottleneck described in why AI made writing faster but publishing slower.
Cross-vendor consensus is the only method that scales. Send the source and the claim to several independent models from different vendors, and ask each one to pull out the supporting passage. Those models are trained on different data, with different alignment, so their citation errors do not line up. Joint agreement is then a much stronger signal than one model's confidence, and disagreement is a flag you can act on. It points you at exactly the citations to read yourself.
This is the shape behind TrueStandard. Paste your draft, and four to five frontier models from different vendors check every claim and citation in parallel. In about 60 seconds you get a replayable log that shows where they agree, and more usefully, where they disagree. It does not remove your judgment on the claims that matter; it removes the 45 minutes of lookups that stood between your draft and that judgment.
Frequently Asked Questions
What is the fastest way to tell if an AI citation is fake?
Open the source and match its title, authors, and year against the reference, character for character. Then find the sentence that supports the claim. Do not stop at the link resolving: in 2026 the common failure is a working identifier attached to the wrong title.
ChatGPT gave me links to its sources. Doesn't that mean they are real?
No. A link can resolve to a real page that is not the cited work, or to a real work that does not contain the claim. A link tells you the model produced a URL, and it does not tell you the URL supports your sentence. Both still need checking.
How common are fake AI citations?
Common enough to be measured at scale: audits of millions of papers show a multi-fold rise in fabricated references since 2023. Library testing has found large shares of AI-generated references wrong or unverifiable, so treat every unchecked AI citation as suspect.
Can an AI detector tell me if my citations are fake?
No. AI detectors guess whether text was machine-written, and they say nothing about whether a citation exists, or whether it supports a claim. Spotting the author and checking the source are different jobs, and only the second keeps a fabricated reference out of print.
Is a reference-checking tool enough on its own?
It is a triage filter, not a sign-off. Tools handle existence well, and some metadata well, but they rarely check that a source supports your claim, and they can flag real citations as fake. Use one to find where to look first, then read the sources for the claims that carry risk.
Can I just ask another AI to check the citations?
A second model from a different vendor helps, because its errors line up less with the first. One model checking its own output is a closed loop. The version that works is several independent models checking at once, and they surface disagreement: that is what cross-vendor checking does.
Keep reading
Why AI Hallucinations Are Structural
DELEGATE 52, GPT-5.5, and a Purdue impossibility proof. Three April 2026 results that move 'hallucinations are structural' from take to documented fact.
Why AI Can't Check Its Own Work
A model carries the same blind spots into review that it had while writing. Dressing it up as a critic is a costume, not a second mind.
What Is AI Sycophancy?
Your AI agrees with you too much. Anthropic's safeguards team explains why models tell you what you want to hear, and what you can do about it.
California's Verify Every AI Output Rule
Three states proposed or enforced 'independent verification' for AI work in 30 days. Here is what 'independent' actually requires.
TrueStandard vs FactCheckTool
These two tools look alike, but they solve opposite problems. One tells you if the media you read is fake. The other tells you if the draft you are about to publish is true.
What we measured on this model
Each release page carries our own fabrication data for one model version, measured against the model it replaced in the same vendor's line.
Catch Fake Citations Before Readers Do
TrueStandard runs your draft through four frontier models from different vendors in parallel, and surfaces every citation they disagree on, in about 60 seconds. A log you can show.
Start Verifying →