Are those AI citations real? Audit them before you trust them.

Paste AI-generated text below — get every URL, DOI, arXiv ID and reference checked for hallucination-risk patterns, with a per-citation verdict, an overall trust score and a downloadable report. Audited entirely on your device.

  • Your text data never leaves the device
  • URLs, DOIs, arXiv IDs, bracketed numbers, references
  • Per-citation verdict + trust score
  • Downloadable audit report

Paste AI-generated text with its citations

Include the reference list for the best results — bracketed numbers and author–year citations can only be matched against entries the tool can see. Audited on the first 200 KB of pasted text.

Try a sample:

Audited in your browser — every citation is checked for format and consistency, and nothing is uploaded.

    No text yet. Everything stays on this device. The report downloads instantly.

    Privacy you can verify, not just trust

    Most citation checkers send your text to their server — the best-known ones resolve every DOI against a database on their side, and some require an account before they check anything. This one can't: your text is audited entirely in your browser. Your file data never leaves the device — the only third-party requests are the site's own analytics, which count page views — and it uploads nothing, ever. There is no server call in the audit at all. Don't take our word for it:

    New cross-origin requests during use: 0
    1. Press F12 to open developer tools, then the Network tab.
    2. Audit a text above. Watch: apart from the site's analytics, nothing new is requested — no uploads, and no file data ever leaves your device.
    3. Extra credit: disconnect your internet after this page loads — auditing still works.

    How to check AI-generated citations without uploading them

    AI models invent references that look real. The authors are often real researchers in the right field. The journal exists. The DOI has the correct format. Only the paper is missing — and in studies of ChatGPT-generated literature reviews, more than half of the references were fabricated. The fix is not to trust the polish; it is to check every citation before the work leaves your desk.

    This tool does the mechanical part of that check where the text already is: on your own device. When you paste an AI answer here, your browser:

    1. Extracts every citation it can find — URLs, DOIs, arXiv IDs, bracketed numbers like [1], author–year citations like (Smith, 2020), and structured reference-list entries under a References heading.
    2. Audits each one for format and consistency — whether the DOI prefix belongs to a recognised publisher, whether the domain is a real scholarly host or a typo-squat, whether the year is in the future, and whether in-text citations have a matching entry in the reference list. Each citation gets a verdict — Sound, Check this, or High risk — with every individual check shown.
    3. Scores the whole answer and writes a report — an overall trust score from 0 to 100, a list of orphans and duplicates, and a downloadable plain-text report you can keep with the draft or hand to a reviewer.

    What makes this auditor different

    The pages that top the results for this search are mostly guides that teach the five manual steps, or paid services that resolve every DOI against Crossref, PubMed and OpenAlex behind an account — the leading dedicated checker is a commercial product. This page is the only free, fully client-side citation auditor we found that shows its work: every citation extracted in the open, every check listed per citation, orphans and duplicates named, a trust score, and a downloadable report — with no account, no daily limit and no upload — plus a live network-request counter so you can verify the privacy claim yourself. It runs entirely in your browser — there is no server to upload to.

    When is auditing the right choice — and when is it the wrong one?

    Auditing is right whenever AI-generated text is about to be submitted, published or relied on: essays, literature reviews, reports, grant drafts and articles that cite sources the model supplied. It is the wrong choice for text with no citations at all — there is nothing to audit — and for a claim whose source you already opened and read, where the audit adds nothing. It is also not a substitute for reading: the auditor catches the mechanical failures, and only a human with the source open can judge whether the source actually supports the sentence beside it.

    Honest limitations

    Format, not verification

    This audit checks whether a citation is well-formed and internally consistent. It does not fetch the URL or query any database, so it cannot prove a source exists.

    A pass is not proof

    A citation can pass every check here and still point at a real paper that does not support the claim beside it. Claim support needs a human reading the source.

    Small publishers can be flagged

    The recognised-publisher list covers the major houses. A real DOI from a small or new publisher is flagged as unrecognised — resolve it at doi.org instead of assuming it is fake.

    Reference parsing is a heuristic

    Reference entries are parsed from author, year, title and journal patterns. Unusual citation styles or heavily abbreviated entries may not parse, and will be reported as missing.

    Sampled, not whole-text

    Auditing is capped at the first 200 KB of pasted text. A very long draft is audited on its opening — paste the rest separately if you need the whole thing checked.

    It never changes your text

    This tool only reports. Your pasted text is never modified — to fix a bad citation, remove it or replace it with a source you found yourself.

    Frequently asked questions

    Does my text upload anywhere?

    No. Your text is audited entirely in your browser — your file data never leaves the device, and the only third-party requests are the site's own analytics, which count page views. You can verify it yourself with the network-request counter in the privacy section below the auditor.

    How does the auditor check a citation?

    It extracts every URL, DOI, arXiv ID, bracketed number like [1], author-year citation like (Smith, 2020) and structured reference-list entry from your pasted text, then checks each one for format and consistency: whether the DOI prefix belongs to a recognised publisher, whether the domain is a real scholarly host or a typo-squat like .con, whether the year is in the future, and whether in-text citations have a matching entry in the reference list. Each citation gets its own verdict — Sound, Check this, or High risk — with the individual checks shown.

    Can this tool prove a citation is real?

    No — and no honest tool can claim that from the citation alone. This audit checks format and internal consistency only: it does not fetch the URL, query Crossref or PubMed, or read the source, because a browser page cannot do live verification of arbitrary sites without hitting cross-site limits. A citation can pass every check here and still point at a real paper that does not support the claim beside it. Treat a Sound verdict as a green light to do the manual checks, not as proof — open the DOI at doi.org, search the exact title, and find the passage that supports the claim.

    What is a hallucinated citation?

    A reference an AI model invented instead of retrieving: the authors are often real researchers in the right field, the journal exists, and the DOI has the right shape — only the paper is missing. Studies of ChatGPT-generated literature reviews found 55% of GPT-3.5 references and 18% of GPT-4 references were fabricated, and later tests found similar patterns across other chatbots. The classic forms are fully fabricated references, real authors attached to a paper they never wrote, DOI strings that resolve to nothing or to a different paper, and real papers cited with the wrong year or details.

    What does the trust score mean?

    The score starts at 100 and loses 25 points per High-risk citation, 10 per citation flagged Check this, and 5 per orphan or duplicate. 85 to 100 means the citations look sound overall, 60 to 84 means the text needs checking, and below 60 means high hallucination risk. It is a triage score, not a certificate: it measures how many format and consistency problems were found, not whether the sources are real or support the claims. Always read the per-citation checks — a single High-risk DOI in an otherwise clean list can matter more than the score suggests.

    Why is a citation flagged when the journal looks real?

    Because a plausible journal name is the easiest part of a citation to fake — language models are good at producing real journal titles, real author names and correctly shaped DOIs. The auditor flags what the journal name cannot prove: a DOI prefix that belongs to no recognised publisher, a domain with a typo-squat ending like .con, a year in the future, or an in-text citation with no matching entry in the reference list. Those are the signals that survive a polished surface, which is why the tool checks them instead of trusting how professional the reference looks.

    What should I do with a High-risk citation?

    Do not keep it in your work until you have checked it by hand. Resolve the DOI at doi.org and confirm it leads to a paper with the same title and authors; if there is no DOI, search the exact title in quotation marks in Crossref, Google Scholar or PubMed; then open the source and find the passage that supports the sentence beside the citation. If nothing matches, remove the citation or replace it with a source you found yourself — never ask the same model to reassure you that its own citation is real.

    Does it check whether the source supports the claim?

    No. Claim support is the hardest check of all and no automated tool can do it from the citation alone — it requires reading the source and comparing it with the sentence it sits beside. A working link proves only that a page exists, and a real paper can still be cited for something it never says. The auditor catches the mechanical failures (invented DOIs, broken domains, orphan numbers, mismatched authors and years) so your manual reading time goes to the one check only a human can do: does this source actually say this?

    What kinds of citations can it audit?

    URLs (http and https), DOIs (bare or as doi.org links), arXiv IDs, bracketed numbers like [1] and [2, 4], author-year citations like (Smith, 2020) and Smith et al. (2020), and structured reference-list entries under a References, Works Cited or Bibliography heading. Paste the full AI answer including its reference list for the best results — bracketed numbers and author-year citations can only be matched against entries the tool can see. Malformed addresses like htps:// or a bare www. address are flagged as typed-from-memory, not copied.