Blog
Research

Hallucinated citations: arXiv's one-year ban, the misconduct debate, and what reference checks miss

arXiv bans authors for a year over unchecked AI references, and a Nature correspondence calls hallucinated citations misconduct. A reference check catches fake papers. It misses real papers cited for claims they never made.

GZ

Guy Zana · Founder

November 9, 2026 · 5 min read

Share
Hallucinated citations: arXiv's one-year ban, the misconduct debate, and what reference checks miss

arXiv's computer science section now bans authors for a year when there is incontrovertible evidence they did not check AI output, and hallucinated references are the named example. A Nature correspondence in October went further and argued that AI-hallucinated citations should be treated as research misconduct.

A reference check catches a paper that does not exist. It cannot catch a real paper cited for a claim it never made.

That difference matters when you prepare a manuscript. An accurate title, author list, and DOI tell readers which paper to find. They do not tell readers whether that paper actually reported the result you attribute to it.

When might a hallucinated citation be misconduct?

An AI tool can produce a plausible reference to a paper that does not exist, or combine real authors, titles, journals, and identifiers into a false entry. Researchers remain responsible for the citations they submit, regardless of how a draft was produced.

In an October 2026 Nature correspondence, Jan Pisek and Eve Veromann argue that AI-hallucinated citations should be treated as research misconduct.

Research integrity scholars David Resnik and Mohammad Hosseini make a narrower argument about US research misconduct regulations. They argue that hallucinated citations may constitute fabrication when the citations function as research data and the author shows indifference to the risk of fabrication.

Their analysis distinguishes misconduct from an honest error. The applicable institution and funding rules matter for an actual case.

A fabricated citation warrants correction. An allegation of misconduct requires a separate assessment of the circumstances. A typo in a DOI, a misread paper, and an invented source can all damage a manuscript, but they do not establish the same facts about an author's conduct.

arXiv's one-year ban for unchecked AI references

arXiv's moderation policy says that authors “each individually take full responsibility for all its contents” and includes incorrect AI-generated references in that responsibility. The policy page does not specify a one-year penalty.

In May 2026, Thomas Dietterich, chair of arXiv's computer science section, described a one-year submission ban for cases with incontrovertible evidence that authors did not check AI-generated output. He named hallucinated references as an example.

He also described a requirement for subsequent submissions to be accepted at a reputable peer-reviewed venue first.

That announcement concerns evidence of unchecked output. Check the current rules of your intended venue and institution before submitting.

What a reference check can establish

A DOI can help you find the work an entry is supposed to identify. Compare the DOI's destination with the title, authors, year, and publication in your reference list. Search the publisher or another reliable scholarly index when an identifier is missing or points elsewhere.

ProblemWhat to look forNext action
A paper appears not to existNo reliable record matches the authors and titleFind the intended source or remove the reference
A DOI points to another workThe destination has a different title or author listFind the correct identifier and check the citation
Details belong to different worksThe title exists but the journal, authors, or year do not matchCompare the entry with the publisher's record
A real paper does not support the sentenceThe relevant passage makes a different or narrower claimRevise the sentence or find appropriate evidence

A failed lookup is a reason to investigate, not proof of fabrication. An index can omit a work, metadata can contain errors, and papers can appear under different versions or titles.

The second question requires the paper itself.

A real paper can still be the wrong evidence

Imagine a paper finds that a reading intervention improved test scores in one classroom. A sentence citing it says the intervention improves reading outcomes across all age groups. Even if the reference is accurate, the sentence extends the finding beyond the study population.

The check is different from looking up the DOI. Compare the sentence with the relevant passage, including the population, result, uncertainty, and context. A paper may also describe another author's claim without endorsing it. Read the surrounding text when attribution or qualification matters.

A source can be relevant without supporting every part of a sentence. If a sentence combines several findings, check which source supports each finding. For a detailed method, see how to check if a citation supports a claim.

Two checks before you submit

If any part of your draft passed through an AI tool, run both checks below before you submit.

Start with the reference list, then check the claims in the manuscript. A reference checker can prioritize entries with missing records or mismatched identifiers. Open flagged records and correct the underlying entries, rather than assuming every flag means a fabricated paper.

Next, compare each substantive claim with the source it cites. Give particular attention to numbers, causal language, populations, and claims summarized from several papers. If the passage does not support the wording, revise the sentence or locate a source that does. Keep a record of the correction for coauthors.

Automated claim checks can help identify passages to inspect. Their judgments depend on the evidence supplied and can misread a source, so use the paper to make the final decision. A correct citation alone cannot establish the quality of a study's methods or conclusions.

Check a claim in Agent Bayes

Agent Bayes keeps cited research claims in a mindmap. Its AI Verification compares a selected claim with its attached passages and reports discrepancies.

In a separate paper review example, AI Support Check searched a library and flagged all seven deliberately planted citation mistakes in a short draft.

To check a claim, select a cited statement in an Agent Bayes mindmap and press F3. Read the discrepancies it adds under the claim, then open the cited passage and compare it with your wording. AI Verification is included on paid plans, starting with Starter. Each check holds an estimate of two credits per claim, then charges for actual usage.

New posts, straight to your inbox

No newsletter fluff, just an email when we publish something new.

Email me when a new post is published on the Agent Bayes blog. You can unsubscribe anytime. We'll first send a confirmation email, and we only use your details for this. See our Privacy Policy.

Enjoyed this? Share it.

GZ

Written by

Guy Zana · Founder

We are researchers and engineers building tools that help people reason over large bodies of literature without losing the thread back to the source.