Do Journals Check References in 2026? What They Actually Do

How journals check references 2026 is no longer an open question. The answer comes from the Lancet/CITADEL audit: 12-fold increase in fabricated citations, four major conferences now desk-rejecting papers with hallucinated references, and updated publisher policies.

One in 277 papers published in early 2026 contained a fabricated citation, a 12-fold increase from 2023’s rate of one in 2,828. The Lancet/CITADEL audit of 2.5 million PubMed Central papers identified 4,046 fabricated references spread across 2,810 peer-reviewed articles, and at the time of the audit, 98.4% of those affected papers had received no publisher action whatsoever [1]. That means nearly 2,800 papers with completely invented references had already entered the permanent literature, many of them feeding into systematic reviews and clinical guidelines.

This isn’t a theoretical problem anymore. Four major machine-learning conferences (ICLR, NeurIPS, ICML, ACL) have implemented automated reference-checking systems and started desk-rejecting papers with confirmed hallucinated citations [2] [3]. Springer Nature and Elsevier have updated their editorial policies to explicitly treat AI-generated hallucinated citations as grounds for desk rejection or post-publication retraction [4] [5]. And the Committee on Publication Ethics (COPE) released updated citation recommendations in September 2026 explicitly endorsing automated reference verification at submission because manual checking “does not scale” in high-volume environments [6].

If you’re submitting to a journal or conference in 2026, this article covers what publishers and conferences actually check, what happens when problems are found, the most dangerous hallucination pattern you need to know about, and a practical four-step verification workflow you can run before submission.

What Journals Actually Check for Reference Verification (As of 2026)

Historically, most journals treated the reference list as a formatting exercise. Editors and production staff checked punctuation, journal names, abbreviations, and style consistency, but actual verification that a cited paper existed was rare, reserved for special cases or not done at all. The assumption was that almost all references were genuine and that errors, when they occurred, were honest and easy to spot [7].

That assumption is now unsafe. The scale and pattern of AI-hallucinated citations make it statistically impossible to rely on the premise that “most references are real.” A Taylor & Francis spokesperson told Retraction Watch that the publisher is actively investing in “technology, specialist staff and processes to catch problematic” citations, and that submissions with concerning references are returned to authors for correction [1]. PLOS is “exploring options for system-wide reference integrity screening,” and Jaypee Journals explicitly stated they are moving toward treating reference verification as a core duty rather than an optional post-review step [7].

The technology to cross-check reference lists against indexing databases (Crossref, PubMed, OpenAlex, Google Scholar) already exists. Tools can automatically flag references with no record, incomplete metadata, or suspicious inconsistencies, and integrate into manuscript management systems so that checks run at the moment of submission [7]. The question now is whether journals will deploy them.

COPE’s September 2026 update is the clearest signal so far. The organization explicitly supports implementation of automated verification systems at submission and states that manual checking alone “does not scale” in high-volume environments [6]. This is a direct call for editorial boards to apply verification frameworks when manuscripts are submitted, rather than waiting until concerns have already been raised.

Review articles are at particularly high risk. The CITADEL audit found that review articles had a fabrication rate 57% higher than other paper types [1]. This matters because fabricated references in reviews feed into systematic reviews, clinical practice guidelines, and point-of-care decisions, creating a cascade of false evidence that can affect patient outcomes [7].

Conference Desk-Rejection Policies (The Most Newsworthy Development)

While journals are still building verification into their workflows, machine-learning conferences moved decisively in 2025–2026. Four major venues now treat hallucinated citations as a desk-rejection trigger: ICLR, NeurIPS, ICML, and ACL [3].

ICLR 2026 implemented an automated system that “automatically extracted references from a given submission and checked them against multiple standard bibliographic databases and a standard web search” [2]. All papers with confirmed hallucinated references were desk rejected, with an appeal channel for erroneously flagged papers. ICLR 2026 received 19,525 submissions; 779 were desk rejected overall, and program chairs stated that hallucinated references “partially explain” the unusually high desk-rejection rate compared to past years [2].

However, the ICLR system had a “significant false positive rate”. For example, it would flag a reference to a paper with a non-English title that authors had translated into English [2]. To compensate, ICLR relied on Area Chairs to perform the first round of human review, program chairs to check all flagged references manually, and ensured that every paper with flagged references was reviewed by at least three humans before desk rejection [2]. This three-round human review process partially explains why the 779 desk rejections were not the end of the story for many submissions.

NeurIPS 2026 updated its Main Track Handbook to acknowledge “many cases of hallucinated citations in literature review, which violates the NeurIPS Code of Conduct” [3]. An Area Chair for NeurIPS 2026 publicly confirmed the practical threshold: “2 or more hallucinated references” leads to desk rejection. In one Area Chair’s batch, 5 of 8 submissions had 2+ hallucinated references and were likely to be desk rejected [3].

ICML lists “hallucinated references” explicitly as grounds for desk rejection of individual submissions on its Peer Review Ethics page [3].

ACL provides the most consequential nuance: program chairs confirmed that more than 100 accepted papers were desk-rejected over hallucinated references using an automated system, with every flag reviewed by program chairs and senior area chairs before rejection [3]. But then ACL walked back the policy mid-cycle. After an executive meeting, ACL announced that “affected papers must go through the normal resubmission process” rather than face the desk-rejection decision as final [3]. This real-world example of a policy being adopted, enforced at scale, and then partially reversed demonstrates both the genuine problem and the false-positive challenges that automated systems introduce.

If you’re using AI tools during manuscript preparation, treating hallucinated citations as a desk-rejection trigger in 2026 is not optional. The data from multiple venues shows that the automated systems are real, they’re catching significant numbers of papers, and they have a non-trivial false-positive rate that every venue has acknowledged [2] [3].

For background on why LLMs invent references in the first place, see our article on ChatGPT fake citations and hallucination statistics.

What Happens When Journals Find Errors

When journals or conferences detect fabricated, corrupted, or unverifiable references, the response depends on the stage of the publication process. The scale of the problem and the role of the hallucinated citation in the paper’s conclusions determine the appropriate action.

Pre-submission (conferences): Desk rejection. All four venues discussed above now desk-reject papers with confirmed hallucinated references. ICLR and NeurIPS provide appeal channels, and ACL initially enforced desk rejection before walking back to a resubmission pathway [2] [3].

Pre-submission (journals): Revision requests. Rather than desk rejection, most journal editors ask authors to replace or correct the problematic citations. Taylor & Francis reported that articles with concerning references are “returned to the author,” and submissions with a high proportion of problematic citations are usually rejected [1].

Under peer review: Editor inquiry or clarification requests. COPE’s assessment framework guides editors to examine four dimensions: the nature of the issue, its location in the manuscript, the category of the anomaly, and the scale of the problem [6].

Accepted but not yet published: Late correction at the proofs stage, often accompanied by an integrity self-correction from the authors [1]. Maxim Topaz, lead author of the CITADEL audit, noted that 91% of problematic papers had only one or two fabricated references, “many of which are likely honest mistakes by authors who used AI tools without verifying the output” [1].

Post-publication: Erratum, correction, or retraction. For papers where hallucinated references “fundamentally compromise” the conclusions (Topaz cited an example where 18 of 30 references appeared fabricated), retraction is the appropriate response [1]. For papers with one or two incidental fabricated references, correction and transparency are more proportionate [1].

Bauchner and Rivara (former editors of JAMA and JAMA Pediatrics) argued that any paper containing a hallucinated reference should be retracted, because “researchers ‘incur responsibility for the entire content of that paper’ when they agree to be authors” [1]. David Resnik from the National Institutes of Health disagrees, arguing the response depends on “the role the citation plays in supporting the results of the study” [1]. This is an active debate in publication ethics, and different publishers will take different positions.

The arXiv cautionary tale: The most extreme sanction so far is arXiv’s one-year ban policy. Authors posting papers with “unambiguous” AI hallucinated citations are banned from arXiv for one year, and after the ban lifts, their papers must first “clear peer review at a reputable venue” before arXiv will re-accept them [8]. This is a rare but consequential example of an enforcement action that can materially affect a researcher’s ability to share their work.

The “Topaz Pattern” Explained

The most dangerous hallucination pattern documented by the CITADEL audit is not a completely fabricated paper. It’s something harder to catch: a real DOI paired with an invented paper title. This pattern accounts for 64% of all AI-fabricated citations and is the single most important concept for authors to understand [6] [9].

Here’s how it works. A large language model generates a citation with a fabricated title. The title sounds plausible and matches the topic. The model also generates a valid DOI number, one that belongs to a completely different, real paper. The real DOI resolves successfully to a legitimate publication in a well-known journal, but that paper has nothing to do with the citation the model claimed. When the author resolves the DOI, it “works.” The DOI checker passes. But the resolved paper is entirely unrelated to what the citation claims.

This is why standard DOI-only verification misses 64% of fabricated citations. Resolving a DOI proves that the identifier exists and points to a real paper. It does not prove that the paper you cited is the one that exists [4]. [9]

Below is a comparison of five citation hallucination patterns, including the Topaz Pattern and four other variants that each require a different detection strategy. This table is the single most useful reference for understanding which hallucination types survive which verification methods.

The Topaz Pattern Detection Guide

This table demonstrates why checking DOI-only is insufficient for verification. The Topaz Pattern alone accounts for 64% of hallucinated citations and passes DOI resolution every time. It is caught instantly by the title-comparison step [6] [9]. For more context on how to independently verify whether a reference is real through database cross-checking, see our guide on DOI verification methods.

What This Means for Your Submission

The situation has changed. In 2026, submitting a manuscript with hallucinated citations (whether generated by AI or copied from other papers without verification) carries real consequences at conferences and is increasingly monitored by journals.

The practical recommendation is straightforward: implement a four-step verification workflow before submission. This is not optional if you use AI writing tools during manuscript preparation, and it’s increasingly advisable even for researchers who don’t use AI, given that fabricated references are spreading through the literature [1] [6] [7].

The Four-Step Verification Workflow

Step 1: Resolve the identifier at its official registry. If the reference includes a DOI, resolve it at doi.org. If it has a PMID, verify it at PubMed. Do not trust the DOI link embedded in the citation. Type it manually into the resolver. A DOI that resolves successfully proves only that the identifier exists; it does not prove that the resolved paper matches your citation [9].

Step 2: Compare the resolved title against the citation claim. This is the step that catches the Topaz Pattern. The 64% of fabricated citations that pass DOI resolution 100% of the time are caught here. If the resolved paper’s title does not match the title in your citation, the reference is fabricated regardless of whether the DOI “works.” This step takes approximately 30 seconds per citation and is mandatory for any paper built with AI assistance [6] [9].

Step 3: Cross-check the title in discipline-specific databases. Verify the reference in at least one independent database beyond the DOI resolver. Use Crossref for general journal articles, PubMed for biomedical topics, the ACL Anthology for computational linguistics, OpenAlex as a broad secondary check, and Google Scholar for very recent or regionally specific papers. Different databases cover different content, and using only one leaves verification gaps [6] [9].

Step 4: Verify retraction status. A paper can be real, correctly cited, and still compromised if it has been retracted. Check Retraction Watch or Crossref metadata for retraction notices. A retracted paper can still be cited, but only knowingly and with the retraction acknowledged [9].

Automated Tools Available in 2026

For papers with 50+ references, the manual workflow above becomes time-consuming. Several automated tools offer first-pass screening:

• Hallucinator: Free, open-source CLI tool that extracts references from PDFs and cross-checks against 10 academic databases (Crossref, arXiv, DBLP, Semantic Scholar, PubMed, OpenAlex, etc.) with 99.4% fabrication detection accuracy [9].
• AccuraCite: Independent benchmark of 918 citations (612 real, 306 synthetic) showed 99.4% fabrication recall, catching the Topaz Pattern, blended citations, and other hard cases [9].
• GPTZero Source Finder: Education-focused screening tool for detecting hallucinated citations in student work, trained on 220M+ scholarly articles [9].
• CheckIfExist: Open-source web-based tool with cascading validation against CrossRef, Semantic Scholar, and OpenAlex [8].

Reference managers like Zotero and Mendeley are useful for organizing citations but provide zero built-in verification. They catalog fake references as easily as real ones, so always verify AI-generated citations before adding them to your reference library [9].

What Reviewers and Editors Expect Now

The role of reviewers has shifted. Reviewers are not expected to verify every reference, but they should spot-check key sources supporting major claims. A simple note such as “I could not locate Reference X, which appears central to the conclusion” already gives editors vital information [7]. At conferences, the gap between reviewer scores for papers with hallucinated references versus clean papers was +0.04 at NeurIPS. That is essentially noise. This means that peer review alone will not catch hallucinated citations. Automated systems do the detection, not human reviewers.

The mistake most authors make in 2026 is checking DOI-only and assuming the citation is verified. A real DOI with a fake title (the Topaz Pattern) passes DOI resolution 100% of the time. Title comparison is mandatory [6] [9]. If you’re unsure about specific citation practices or need guidance on building verified reference lists, our FAQ covers common citation-related questions.

The tradeoff here is real: automated verification catches obvious fabrications but introduces false positives (ICLR’s system flagged references to non-English papers, very recent papers not yet indexed, and genuine unpublished work). The human review rounds ICLR implemented were expensive but necessary [2]. Running automated checks yourself before submission is cheaper than being desk-rejected.

If you’re navigating reference verification, conference desk-rejection policies, or the broader scope of AI-generated citations, this article synthesizes data from the Lancet/CITADEL audit (4,046 fabricated citations across 2.5 million papers) with concrete policies from ICLR, NeurIPS, ICML, and ACL. The field is shifting rapidly: publishers are integrating verification tools, conferences are desk-rejecting papers, and COPE is explicitly recommending automated screening. The most practical takeaway is the four-step verification workflow above. Resolve the identifier, compare the resolved title, cross-check in independent databases, and verify retraction status. Run it before you submit.