An AI-invented citation leaves a claim without the source it appears to have. Existing policies give authors concrete responsibilities for their references. Whether a particular invented citation constitutes formal research misconduct is a separate question that this explainer does not resolve.

The sources have different scopes: arXiv sets platform rules, the International Committee of Medical Journal Editors (ICMJE) makes medical-journal recommendations, and the US Office of Research Integrity (ORI) publishes a definition in its Public Health Service and Department of Health and Human Services context.

The immediate prompt is a Nature correspondence by Jan Pisek and Eve Veromann, published on 27 October 2026. Its title argues that “AI-hallucinated citations should be treated as research misconduct.” That is the authors’ advocacy. The full correspondence is access-limited, so its argument is not reconstructed here.

An invented source and a misleading citation need different checks

“Hallucinated references” commonly describes citations generated by AI that appear to identify scholarly work but contain invented bibliographic information. The practical problem includes two distinct possibilities: the cited work does not exist, or a real work does not support the statement attached to it.

Finding a matching paper addresses the first question. Reading the relevant passage addresses the second. A reference can have the correct title, authors and identifier while still being unsuitable evidence for the claim.

Four questions help keep the issue precise:

  • Is the reference accurate?
  • Has the source been checked for the statement it supposedly supports?
  • What obligations apply under the relevant platform or journal policy?
  • Has formal research misconduct been established?

These questions can overlap. They are not stages in an escalation process, and a faulty reference alone does not explain how or why it entered a manuscript.

arXiv makes authors responsible for incorrect AI-generated references

In its policy on authors’ use of generative AI language tools, arXiv says each named author takes full responsibility for the paper’s contents, regardless of how those contents were generated. The examples explicitly include “incorrect references.” Its scholarly standards also call for carefully prepared references.

That gives authors a direct obligation: using AI to produce part of a manuscript does not transfer responsibility for its references to the tool.

arXiv also has general moderation and conduct measures. Its moderation policy allows it to decline submissions and describes possible action on policy violations after announcement, including withdrawal or a concern recorded in the paper’s metadata. Its code of conduct says violations may result in a warning or suspension from participation.

These provisions establish responsibility and discretionary measures. They do not establish an automatic citation-specific rejection or account ban. Platform moderation also answers a different question from a formal institutional or federal misconduct finding. The policy passages discussed here were checked on 6 November 2026.

ORI’s definition does not settle the classification of invented references

The US Office of Research Integrity’s definition webpage defines research misconduct as fabrication, falsification or plagiarism in proposing, performing or reviewing research, or reporting research results. It defines fabrication as “making up data or results and recording or reporting them,” and excludes honest error and differences of opinion.

That wording requires care when discussing a “fabricated citation.” Calling a bibliographic reference fabricated in ordinary language does not, by itself, establish that it meets the regulatory meaning of fabrication.

The webpage does not explicitly classify invented bibliographic references, and it is not the complete operative rule. This article therefore leaves that formal classification unresolved. Nor does a bad reference alone establish that an error was honest.

ICMJE recommends checking accuracy and support for the claim

The ICMJE recommendations on references, in section IV.A.3.g.i, place responsibility for citation accuracy on authors. Authors should be able to attest that references support their associated statements, and should verify references against an original source or a bibliographic source such as PubMed.

ICMJE also recommends checking that references do not cite retracted articles, except when discussing the retraction. For articles in MEDLINE-indexed journals, it identifies PubMed as authoritative for retraction information.

These are medical-journal recommendations. Their useful practical distinction is between confirming a reference’s identity and confirming what it supports.

A compact way to check an AI-generated reference

The following procedure puts ICMJE’s verification and claim-support recommendations into practical steps. The metadata checklist, correction and version checks, and handling of unresolved references are editorial synthesis.

  1. Find the source record. Locate the original publisher or repository page, or a reliable bibliographic record. Compare the title, authors, year, journal or repository, and identifier. If a DOI is supplied, inspect the record it resolves to. A resolving DOI can identify a work; it does not establish that the work supports the claim or is reliable.

  2. Read for the specific claim. Find the relevant passage and read its context. Check whether it supports the statement as written, including any conditions or limitations. A real source can still be cited misleadingly.

  3. Check its current status. Look for later versions and linked correction or retraction notices. Read what each notice changes before deciding how it affects the claim. Nature’s 30 October correction to the correspondence discussed above, for example, restored Veromann’s omitted name and affiliation.

  4. Resolve mismatches before using the reference. A failed lookup is a reason to investigate spelling, incomplete metadata, alternative versions and access routes. It alone does not prove invention. If the source’s identity or support for the claim remains unverified, omit or replace the reference and revise the claim accordingly.