The AI Fabrication Shield is our name for the human verification process that checks every citation, statistic, and factual claim in your manuscript against its original source before the work is delivered. This page explains the problem it addresses, exactly what we check and how, an honest account of what it can and cannot catch, and what you receive.

The problem, in primary evidence

Large language models frequently produce references that look authoritative but are fabricated, misattributed, or drawn from retracted work. This is documented in the peer-reviewed literature, not speculation.

In a controlled study published in the Journal of Medical Internet Research, Chelli and colleagues tested ChatGPT and Bard against gold-standard systematic reviews and measured how often each model produced hallucinated references. The study reported substantial hallucination rates and concluded that these tools cannot yet be relied upon to generate accurate references for scientific writing (Chelli et al., 2024, JMIR, doi:10.2196/53164).

A 2025 experimental study by Linardon and colleagues, also in JMIR Mental Health, prompted GPT-4o to generate literature reviews and examined the resulting citations. It found that a meaningful proportion of references were fabricated — unable to be traced to any real publication — and that fabrication rates rose for less-established topics. Critically, the authors noted that fabricated citations often carried real-looking details, making them difficult to detect without deliberate source checking (Linardon et al., 2025, JMIR Mental Health, doi:10.2196/80371).

The practical lesson from this body of work is consistent: a fabricated reference is often indistinguishable from a real one on the page. The only reliable defence is to check each citation against the primary source — the original article on the journal or publisher site — rather than trusting a generated reference at face value. That is what the Shield does.

What can go wrong with an unchecked citation

The failure modes we screen for are specific:

  • The reference does not exist. A plausible author, title, journal, and year that correspond to no real paper.
  • The DOI points somewhere else. A real, resolvable DOI that links to a completely different paper than the one named.
  • The claim is not in the source. A real paper cited for a finding it never reported.
  • The source has been retracted. A real, correctly cited paper that has since been withdrawn and should not be relied upon.

Any one of these can lead to desk rejection, reviewer distrust, a post-publication correction, or — in the worst case — a retraction recorded against your name.

The four verification stages

Before delivery, a human editor takes every reference and claim through four checks:

Stage 1 — Existence check. Each DOI and PMID is resolved to its source of record. We confirm the reference against the original journal or publisher page — the article of record itself — and cross-check it through indexing services such as Crossref, PubMed, and the publisher’s own database. The author, title, journal, volume, and year must all match the record. A DOI that resolves to a different paper is treated as a failure, not a pass.

Stage 2 — Claim check. The specific statement attributed to each source is read against that paper’s abstract or, where access allows, its full text on the original journal site. If the source does not actually support the claim, the claim is rewritten to match what the evidence says, or the citation is replaced with one that does support it.

Stage 3 — Retraction check. Each cited paper is screened against the Retraction Watch Database and PubMed retraction notices, so withdrawn work does not silently enter your reference list.

Stage 4 — Expert read. A team member with relevant doctoral or clinical expertise reviews the cited evidence in context, confirming that the references genuinely support the argument being made — not just that they exist.

A worked example

Suppose a draft contains this sentence: “Intervention X reduced recurrence by 38% in a randomised trial (Author A et al., 2021).”

  • Stage 1 finds the DOI resolves — but to a review article, not a randomised trial, and the author list does not match. Flagged.
  • Stage 2 confirms no 38% recurrence figure appears anywhere in the resolved paper. The claim is unsupported.
  • Resolution: the editor either locates the genuine primary trial that reports the effect and corrects the citation and the figure to match it, or, if no such source exists, removes the unsupported claim. The sentence that reaches the journal is one whose number and citation are both real and aligned.

This is the kind of error that reads perfectly on the page and passes casual review — which is exactly why a deliberate, stage-by-stage human check is the only thing that reliably catches it.

What the Shield does not claim

We are precise about scope, because overstating it would be its own form of the problem we are trying to solve. The Shield verifies that references exist, that claims match their sources, and that cited work is not retracted at the time of delivery. It does not, and cannot, guarantee that a reviewer will agree with your interpretation of a source, that a paper will not be retracted in the future, or that every judgement call in a complex literature is the only defensible one. What it guarantees is that no citation in your manuscript is fabricated, misattributed, or knowingly retracted when we hand it back.

How this fits our wider approach

The Shield reflects a single principle that runs through all of our work: where any AI assistance is used, it is always placed under human supervision and grounded in primary sources, and it is never left unchecked. Our full position on this is set out in our AI Use Policy.

What you receive

  • A manuscript in which every citation has been verified against its original source — the journal or publisher article of record — not just an index entry.
  • References screened for retraction before delivery.
  • Claims that match the sources they cite.
  • A reference list you can stand behind in review and after publication.

The Shield is included on every manuscript we write or rewrite. To discuss your project, contact our team.

References

Chelli M, Descamps J, Lavoué V, et al. Hallucination Rates and Reference Accuracy of ChatGPT and Bard for Systematic Reviews: Comparative Analysis. J Med Internet Res. 2024;26:e53164. doi:10.2196/53164

Linardon J, Jarman HK, McClure Z, et al. Influence of Topic Familiarity and Prompt Specificity on Citation Fabrication in Mental Health Research Using Large Language Models: Experimental Study. JMIR Ment Health. 2025;12:e80371. doi:10.2196/80371