Skip to content
Citations Citations
Blog About Request Access

AI Citation Integrity in Medical Publishing: From Plausible References to Verified Sources

Francois-Xavier Bioul
Francois-Xavier Bioul · CCO at Citations LLC
10 min read

A citation can look real and point to nothing. It can carry a journal name, a plausible author list, a formatted DOI, a date and a title that fits the surrounding claim, and still be invented.

For medical publishers, this has become a structural threat to confidence in the source, which is the asset medical publishing depends on. The standard publishers need is whether a cited source exists, matches trusted metadata and was used in a traceable workflow, beyond whether an AI-generated citation looks plausible.

In short

AI citation integrity in medical publishing concerns whether a cited source can be verified before it enters the published record, beyond whether the reference is correctly formatted.

A fabricated citation may include a plausible title, journal, author list, date or DOI-like string. If the source does not exist, cannot be resolved, or does not match trusted metadata, the citation contaminates the record instead of supporting it.

The 2026 Lancet audit of biomedical literature shows that the problem is now structural. Fabricated references are passing through AI-assisted writing, submission, review and publication workflows because many systems still treat references as text rather than as verifiable source events.

At Citations, we use the term usage evidence to describe rights-aware records that make AI content use visible, attributable and commercially actionable. For medical publishers, the principle is that a citation should be verified, matched to a trusted source record and logged as part of the workflow.

What the Lancet audit found

In May 2026, The Lancet published an audit of 2.5 million biomedical papers from PubMed Central’s Open Access subset. Among papers published in early 2026, approximately one in 277 contained at least one fabricated reference. In 2023, the rate was approximately one in 2,828.

Columbia researchers identified 4,046 fabricated citations across 2,810 papers. Many were correctly formatted and on topic, which is why they went unnoticed: a fabricated reference that looks wrong is usually caught, while one that looks right can circulate.

For this reason, AI usage evidence for publishers cannot remain a licensing-only discussion. In medical publishing, evidence also covers source integrity.

Publisher action after detection

The audit also reported that, at the time of analysis, 98.4% of affected papers had received no publisher action: no correction, no retraction and no visible integrity signal. That turns a research integrity finding into a publishing infrastructure problem.

Generative AI can produce fabricated references, and many publishing workflows still treat a reference as a string of text rather than as a verifiable source event. A reference that renders correctly and fits the bibliography may pass without anyone checking whether the source exists, whether the metadata matches, whether the DOI or PMID resolves, whether the source is accessible, whether it has been corrected or retracted, or whether it was actually used in the AI-assisted output.

That gap was easier to tolerate when fabricated references were rare, produced by hand and easier to spot. It becomes a serious risk when generative AI can produce plausible references at scale.

Why AI raises the stakes

Fabricated citations existed before AI. What AI changes is the speed, volume and surface area of the risk. A model can produce a reference that passes the visual tests of legitimacy without any underlying source, and AI-assisted writing, summarization, literature review and manuscript preparation increasingly sit upstream of the published record.

A fabricated reference can therefore be created inside an AI workflow and enter the literature through normal editorial channels. Once published, it behaves like any other citation: it can be retrieved, summarized, used to support a downstream claim, cited in a review, used in a guideline and absorbed by another AI system. The error compounds, and no step in the chain necessarily checks whether the source ever existed.

Machine-readable provenance for reference publishers addresses this as an operational control. Medical publishers need source records that systems can check before fabricated references move downstream.

Why this is a medical publishing problem

A weak citation in a consumer article is a quality problem. In medical literature, it can become an evidence problem. A clinician does not verify every reference behind every review, a guideline developer cannot reconstruct every citation trail, and a medical publisher cannot protect trust if the first warning arrives after an article has entered downstream evidence workflows.

Fabricated references are therefore a workflow verification issue as well as an author misconduct issue. A publisher’s authority rests on the trust signals around the text: peer review, editorial governance, version control, corrections, retractions, metadata quality and provenance. Each citation that cannot be traced to a real, accessible, correctly versioned source erodes that authority, one paper and one dataset at a time.

From plausible references to verifiable source events

The traditional check asks whether a citation looks correct. The check medical publishing now needs asks whether the cited source can be verified. A reference is a claim about a source, and that claim should be checked against trusted source records, including:

  • DOI or PMID;

  • title;

  • authors;

  • journal;

  • publication date;

  • version;

  • publisher record;

  • correction status;

  • retraction status;

  • source accessibility.

The check should take place when the citation enters the AI-assisted workflow, before submission, editorial review, publication and downstream reuse. Once a fabricated reference reaches the published record, every downstream system inherits it.

Citation as text vs citation as evidence

The two approaches differ on every point that matters to a medical publisher.

Question

Citation as text

Citation as evidence

What is checked?

Formatting and plausibility

Source existence, metadata, accessibility and integrity status

When is it checked?

Often late, manually or inconsistently

At the point of workflow entry

What does it prove?

The reference looks acceptable

The source can be verified

Main risk

Plausible hallucinations pass through

Requires connected source records and audit events

Publishing value

Editorial appearance

Source confidence

A formatted citation, a plausible reference or a DOI-like string does not establish that the source exists. Medical publishers need verification for that.

What a connected publishing catalogue makes visible

Policing every author manually does not scale. A more practical approach is to make the source layer observable at the point of AI use. When a publisher’s catalogue is connected to the AI workflow, each claimed citation can be checked against trusted source records, and each AI-assisted output can generate an audit event.

That event links the claimed citation, the matched source record, the result of the metadata check, the integrity status and the workflow decision, which may be to allow the citation, flag it, block it or send it for human review.

This would not eliminate every fraud pattern, and determined misconduct will keep evolving. It would make many phantom citations visible before they pass as formatted, plausible text. A reference with no matching record should not move through a medical publishing workflow unnoticed.

Retrieval, citation and verification are separate events

AI search adds a further complication. A source can be retrieved without being cited, which creates a visibility problem for publishers. In medical publishing the reverse also matters: a source can be cited without being verified, which creates an integrity problem.

Both failures come from the same weakness, a workflow that does not keep enough evidence about source use. That is why retrieved is not cited belongs in the same evidence chain. AI systems separate retrieval, citation, verification, usage, attribution and source existence, and none of these should be inferred from the others. A citation does not prove that a source was used, a retrieval event does not prove that it was cited, and a formatted reference does not prove that it exists. Each event needs its own record.

Provenance and usage evidence

Provenance helps establish where content came from and how it has changed over time. Citation integrity also needs usage evidence, because an authentic source can still be misused. A source may exist and not support the claim, or a citation may point to a real article while mismatching its title, authors, date, version or conclusion.

The distinction in content provenance vs usage evidence is therefore editorial as well as commercial. Provenance answers whether the source is real and traceable. Usage evidence answers whether it was used in this workflow, for this output, under these conditions. Medical publishing needs both.

What source verification should capture

At minimum, a source verification event in an AI-assisted medical publishing workflow should capture:

  • the claimed citation;

  • the matched source record;

  • the DOI, PMID or other identifier;

  • metadata consistency;

  • source accessibility;

  • publication status;

  • version status;

  • correction or retraction status;

  • the AI workflow in which the source was used;

  • the decision taken by the system;

  • the human review status, if needed;

  • the date and time of verification.

This record turns citation checking into an auditable workflow control instead of a post-publication clean-up. Without it, publishers depend on visual plausibility, which generative AI reproduces easily.

The same gap appears in editorial policy: AI disclosure is not an audit trail for STM publishers, and citations are where disclosure fails first.

Why proof of content use matters here too

Fabricated citations and AI content licensing look like separate problems, but both depend on records of source use. In licensing, publishers need to know which content AI systems used, to support pricing, attribution and renewal. In medical publishing, they need to know which sources were used or claimed inside AI-assisted workflows, to protect the integrity of the record. The commercial consequences differ, and the evidence problem is the same.

For that reason, proof of AI content usage is relevant here: a medical publisher cannot protect value or trust without verifying source use.

The objection: reference checking already exists

Reference checking tools already exist. Editorial teams use DOI checks, Crossref tools, PubMed links, manuscript systems and production workflows.

The Lancet audit shows that these processes are not sufficient. The gap lies between available verification and workflow-level enforcement: a fabricated reference can still pass if verification is optional, late, manual, incomplete or disconnected from AI-assisted drafting. Medical publishers need source verification embedded in the workflow, so that a citation that fails verification becomes a logged workflow event that is flagged, blocked or reviewed, and can be audited later.

What Citations is built to document

Citations is built on the premise that, in AI-assisted publishing, knowing that a source was mentioned is insufficient. Organizations need to know whether the source existed, was accessible, matched the cited metadata and was used in a traceable way at the point of AI use.

Citations is designed to document source verification, accessibility, usage and workflow traceability. For medical publishers, trust has to rest on verifiable source records and auditable AI-assisted workflows, and source verification has to come before any confidence in the citation.

The question every medical publisher should ask now

Before fabricated references become a recurring item in correction workflows, medical publishers should check whether their team can prove which cited sources were verified before entering an AI-assisted workflow: matched to a real source record, checked for metadata consistency, linked to an accessible version and logged.

If the answer is no, the gap is strategic as well as editorial, because confidence in the source is what medical publishers sell, and an unverified citation undermines it.

Frequently asked questions

What are fabricated citations in medical publishing?

Fabricated citations are references that appear legitimate but point to sources that do not exist, cannot be verified, or do not match the cited metadata. They may include plausible titles, authors, journals, dates and identifiers, which makes them difficult to detect through visual review alone.

Why are fabricated references an AI citation integrity problem?

Generative AI can produce plausible-looking references without grounding them in real source records. The risk increases when AI-assisted writing, summarization or literature review tools are used upstream of submission, review or publication workflows.

How common are fabricated citations in biomedical literature?

A 2026 audit published in The Lancet examined 2.5 million biomedical papers from PubMed Central’s Open Access subset. It found that approximately one in 277 papers published in early 2026 contained at least one fabricated reference, compared with approximately one in 2,828 in 2023.

Why do fabricated citations pass through publishing workflows?

Many workflows still treat references as formatted text rather than verifiable source events. If a citation looks plausible and follows a recognizable format, it may pass without being checked against DOI, PMID, publisher metadata, version status, accessibility, or correction and retraction records.

How can medical publishers reduce fabricated citation risk?

Medical publishers can reduce risk by verifying claimed citations against trusted source registries before publication and before downstream reuse. A source verification workflow should check identifiers, metadata, publication status, accessibility, versioning and integrity signals.

Is source verification enough to eliminate citation fraud?

No. Source verification will not eliminate every fraud pattern. It can make phantom citations visible earlier, especially references that have no matching source record or inconsistent metadata, and it turns citation checking from a manual visual process into an auditable workflow control.

Continue the evidence chain

Start with the broader framework: AI Usage Evidence for Publishers

Make authority machine-operable: Machine-Readable Provenance for Reference Publishers

Separate retrieval from citation: Retrieved Is Not Cited

See where disclosure stops short: AI Disclosure Is Not an Audit Trail for STM Publishers

Distinguish origin from downstream value: Content Provenance vs Usage Evidence

Understand the licensing shift: Proof of AI Content Usage

Ready to assess your evidence layer? Book an AI usage evidence assessment

Sources

The Lancet — “Fabricated citations: an audit across 2.5 million biomedical papers”
Fabricated citations: an audit across 2.5 million biomedical papers

Columbia University School of Nursing — “Nearly 3,000 peer-reviewed medical papers have fake citations, a Columbia Nursing AI-assisted audit finds”
Nearly 3,000 peer-reviewed medical papers have fake citations

Nature — “Surge in fake citations uncovered by audit of 2.5 million biomedical-science papers”
Surge in fake citations uncovered by audit of 2.5 million biomedical-science papers

PubMed — Nature article record: “Surge in fake citations uncovered by audit of 2.5 million biomedical-science papers”
PubMed article record

Retraction Watch — “One in 277 PubMed-indexed papers in 2026 shows fabricated references, says analysis”
One in 277 PubMed-indexed papers in 2026 shows fabricated references