What Should You Do if AI Invented a Citation That Is Already in a Research Draft?

What Should You Do When an AI-Generated Citation in Your Manuscript Does Not Exist?

You are preparing a research manuscript for submission when you attempt to retrieve one of the articles cited in your literature review. The reference includes an author, publication year, journal title, volume, page numbers, and DOI. Everything looks credible.

But the article cannot be found. The DOI may lead to an unrelated publication, or the supposed journal article may not exist at all.

You remember that a generative AI system suggested the reference during an earlier literature review.

Should you simply delete the citation? Can you replace it with another source? Does the presence of one fabricated reference mean the entire literature review must be rewritten?

The immediate problem is bibliographic accuracy, but the larger concern is evidentiary integrity. A nonexistent reference may have been used to justify a claim, establish a research gap, or support a methodological decision. Removing the citation without examining its role could leave an unsupported argument in the manuscript.

Verify the Reference, Reassess the Claim, and Correct the Draft

How to Identify and Correct AI-Fabricated Academic References

What Is an AI-Fabricated Citation?

An AI-fabricated citation is a bibliographic reference generated by an AI system that falsely represents a publication or its identifying details.

Generative AI can combine plausible author names, article titles, journal names, publication years, and DOI-like strings into references that resemble genuine scholarly records.

Some references are entirely nonexistent. Others combine real bibliographic elements incorrectly, such as assigning the title of one article to the authors of another.

A third category involves genuine publications that are cited inaccurately or described as supporting claims they do not establish.

These problems should be distinguished because the corrective action may differ.

Researchers should not automatically describe every incorrect reference as a fabricated publication. The distinction between nonexistent work, inaccurate metadata, and unsupported attribution matters for both correction and research-integrity assessment.

Why Do AI Systems Generate Plausible but Nonexistent References?

Generative language models can produce bibliographic text that follows familiar academic patterns without reliably retrieving or verifying an underlying publication.

A reference may therefore contain convincing combinations of author names, article titles, journal information, and DOI syntax.

The National Institute of Standards and Technology's Generative Artificial Intelligence Profile identifies confabulation as a relevant generative AI risk, including false or unsupported content presented with apparent confidence.

Reference fabrication is one possible manifestation of that broader problem.

The model's confidence does not establish that the publication exists. Nor does a correctly formatted DOI establish that it identifies the stated work.

Even when an AI system has access to retrieval tools, researchers should verify that the retrieved record matches the reference and supports the claim.

First, Verify Whether the Publication Actually Exists

Begin by searching for the exact article title and author combination through reliable bibliographic sources.

Useful starting points include the publisher's website, the journal's official archive, Crossref metadata search, and discipline-specific bibliographic databases.

Google Scholar may help locate alternate versions or records, but search results should be checked against the original publication wherever possible.

If a DOI is provided, resolve it through the DOI system and compare the destination's title, authors, journal, and publication details with the generated citation.

Crossref provides metadata retrieval services that can assist with checking registered publication information.

However, Crossref coverage is not universal. Failure to find a record there does not conclusively establish that a publication is nonexistent.

Older publications, books, conference materials, institutional reports, and certain regional publications may require different verification routes.

The appropriate conclusion is that a source should not be used as evidence until its identity and relevant content can be established reliably.

Do Not Confuse an Incorrect DOI With a Nonexistent Article

A DOI may be wrong even when the intended publication is genuine.

For example, an AI system may provide an authentic article title but attach a DOI belonging to another paper.

In that situation, deleting the entire citation without further investigation could remove a legitimate source.

Search for the publication independently using its title, authors, and journal information. If the original work is found, reconstruct the reference using verified metadata.

Conversely, a DOI that resolves successfully does not prove that the AI-generated title and findings are correct.

The DOI record and the actual publication must match the cited work.

Researchers should verify the publication as a whole rather than treating any single bibliographic field as conclusive.

Examine the Claim That the Fabricated Citation Was Supporting

The most consequential step is determining why the reference appeared in the manuscript.

Suppose an AI-generated citation was used to support the statement that a particular educational intervention consistently improves student achievement.

If the publication does not exist, the statement has lost its purported supporting evidence.

The researcher should examine whether other independently verified research supports the claim, whether the wording needs qualification, or whether the statement should be removed.

A fabricated reference may have influenced more than one sentence. It could have shaped the theoretical framework, research gap, hypotheses, or discussion.

The investigation should therefore trace the citation's role throughout the manuscript.

This is particularly important when the nonexistent publication was presented as a foundational study or as evidence that a research problem had not previously been investigated.

Why Replacing a Fake Citation With Any Real Article Is Not Enough

Finding a genuine publication on a similar topic does not establish that it supports the original claim.

For example, a fabricated citation may supposedly demonstrate that AI tutoring improves examination performance across several countries.

A genuine article about students' perceptions of AI tutoring would not automatically support that effectiveness claim.

Researchers must examine the replacement publication's research question, methods, findings, and limitations.

If the genuine evidence supports only a narrower conclusion, revise the manuscript accordingly.

Replacing nonexistent references with superficially related publications without checking their relevance can create a different integrity problem: genuine sources are used to support claims they do not establish.

The objective is not merely to make the bibliography look authentic. It is to ensure that the scientific argument rests on evidence that actually exists and is represented accurately.

Audit Other References Generated During the Same AI Workflow

Discovering one fabricated citation provides a reason to examine other references produced through the same unverified process.

It does not establish that every AI-suggested source is false.

However, if the same system generated several references without reliable retrieval or verification, those records may share similar risks.

Identify which references were supplied by AI and whether they were independently checked before incorporation.

Prioritize sources supporting consequential claims, including the study's rationale, theoretical framework, methodological decisions, and principal interpretations.

A complete reference audit may be appropriate when numerous citations were generated through the same workflow and their verification status is uncertain.

Researchers should avoid relying on AI to certify the authenticity of the very bibliography it generated.

How to Conduct a Citation Verification Audit

A practical audit separates bibliographic existence from evidentiary relevance.

For large bibliographies, reference-management software and bibliographic metadata services can assist with organization and duplicate detection.

Nevertheless, automated metadata matching cannot establish that a source supports a particular interpretation. Claim-level verification remains a scholarly responsibility.

What if the Fabricated Citation Supports a Central Research Claim?

The seriousness of a fabricated citation depends partly on its role in the manuscript.

A nonexistent reference attached to a minor background statement may require a limited revision if the statement can be independently supported.

A fabricated citation used to establish the study's central research gap may have more extensive implications.

For example, a researcher may claim that no previous study has examined a particular intervention because an AI-generated literature summary inaccurately represented the field.

Removing one false reference would not resolve the problem if the research gap itself was developed from unreliable information.

The researcher may need to repeat relevant literature searches, reassess the state of knowledge, and revise the rationale for the study.

Where a fabricated citation materially influenced a methodological decision, that decision should also be independently reexamined.

Should the Researcher Disclose That AI Generated the False Reference?

Researchers should distinguish internal documentation, formal AI-use disclosure, and reporting an error to a journal.

The International Committee of Medical Journal Editors recommends disclosure of AI-assisted technologies used in producing submitted scholarly work and emphasizes that human authors remain responsible for accuracy and appropriate attribution.

Other journals and institutions may have different requirements.

If the fabricated reference is discovered and corrected before submission, researchers should follow applicable AI disclosure policies and retain relevant internal records where appropriate.

A correction made before submission does not automatically require a separate declaration that a fabricated citation once appeared in a working draft.

However, the original AI contribution may still fall within the journal's disclosure requirements, depending on what the AI actually did.

Researchers should not conceal a consequential unresolved problem or falsely state that references were verified when they were not.

What if the Manuscript Has Already Been Submitted?

A fabricated citation discovered after submission requires assessment of its significance and the manuscript's editorial status.

If the error affects a consequential claim, the authors should consult the responsible coauthors and follow the journal's procedures for communicating corrections.

Possible actions may include supplying corrected manuscript files, explaining changes to the supporting evidence, or revising an affected argument.

Researchers should not assume that an error can safely remain until the next revision opportunity when it materially affects the manuscript under review.

The appropriate response to an AI-generated error discovered after submission depends on the nature of the error and the journal's requirements.

What if the Fabricated Citation Is Already Published?

Once a fabricated reference enters the published record, the issue may require editorial correction.

The appropriate action depends on whether the reference is isolated, whether the underlying claim remains supportable, and whether the problem materially affects the article's reliability.

Committee on Publication Ethics guidance distinguishes corrections from more substantial editorial actions, including retraction when published findings are sufficiently unreliable.

A fabricated reference does not automatically require retraction. Conversely, a collection of nonexistent sources supporting central conclusions may raise serious concerns about the reliability of the article.

Researchers should promptly consult the journal when a verified problem affects published material.

The decision about whether an AI error requires correction of the research record should reflect its consequences and the applicable editorial process.

Does a Fabricated AI Citation Automatically Constitute Research Misconduct?

Not automatically. The presence of a nonexistent reference establishes a bibliographic problem, but a formal misconduct determination depends on the applicable definition and circumstances.

An unrecognized AI hallucination, deliberate fabrication of supporting evidence, and intentional concealment after discovering the problem involve different questions of conduct and responsibility.

Research misconduct rules vary across jurisdictions and institutions.

Researchers should avoid making unsupported conclusions about intent merely because a fabricated reference appears in a manuscript.

Nevertheless, responsibility for checking citations remains with the human researchers who incorporate and submit them.

Prompt correction and transparent handling of consequential errors are important regardless of how the reference originated.

How Can Researchers Prevent AI Citation Hallucinations?

The most reliable preventive practice is to construct the bibliography from independently identified and verified sources.

Researchers can use scholarly databases, publisher websites, reference managers, and DOI metadata services to retrieve authentic bibliographic records.

AI may assist with organizing literature, suggesting search concepts, or identifying topics requiring further investigation. Its generated references should not be accepted without checking the original works.

For each consequential citation, verify both the bibliographic record and the claim-source relationship.

Maintain a clear distinction between sources actually retrieved and read, sources identified for possible review, and references generated by AI but not yet verified.

This reduces the risk of fabricated citations entering the manuscript through repeated copying or automated drafting.

Correcting a Fabricated Citation in an Educational Technology Literature Review

The appropriate correction is not simply replacing the nonexistent article with any authentic publication about AI tutoring.

The original claim depended on specific evidence about effectiveness, scale, and research design. A replacement source must actually support the revised statement.

The example illustrates why fabricated citations can create problems beyond bibliography formatting. They may influence how researchers characterize existing evidence and justify the significance of their work.

Common Mistakes When Correcting AI-Fabricated Citations

A Practical Decision Framework for Fixing AI-Generated References

When a suspicious AI-generated reference appears in a research draft, begin by establishing whether the publication exists and whether its details are accurate.

Then examine the scientific claim attached to it.

Do not treat reference correction as a purely bibliographic task. The goal is to restore the evidentiary integrity of the manuscript.

For a working draft, the problem may be resolved through careful verification and revision before submission.

For a submitted or published manuscript, the same citation error may require additional communication because others are already relying on the affected research record.

The decisive question is whether the corrected manuscript accurately represents the evidence, not merely whether every bibliography entry now looks legitimate.

What to Check After Finding a Fabricated AI Citation

Frequently Asked Questions About AI-Fabricated Citations

Correct the Evidence Behind the Claim, Not Just the Reference List

How to Cite This Guide

This guide is intended to be read, shared, and used in research, teaching, and academic work. If you draw on its ideas, explanations, or other content, please acknowledge the source by citing the guide. Doing so gives appropriate credit and helps your readers locate the original resource.