Can Generative AI Invent Fake DOIs for Research Papers?

If AI Gives You a DOI, Does That Mean the Paper Must Be Real?

A generative AI system recommends a research paper and supplies a DOI. That can make the reference feel substantially safer. After all, a Digital Object Identifier is designed to provide a persistent way of identifying scholarly and other digital objects.

There is one problem: generative AI can generate strings that look like DOIs.

The identifier may not exist. It may exist but belong to a completely different publication. The paper itself may be real while the DOI is wrong, or the entire citation may be fabricated and accompanied by an equally fabricated identifier.

For researchers, therefore, the presence of a DOI is not verification. The DOI itself must be verified.

Can AI Generate a DOI That Is Not Real?

How Can an AI-Generated DOI Be Wrong?

A DOI Is an Identifier, Not Just a Citation-Looking String

A DOI, or Digital Object Identifier, is a persistent identifier assigned to an object through the DOI system. In scholarly publishing, DOIs are commonly associated with journal articles, datasets, books, chapters, preprints, and other research outputs, although not every scholarly work has one.

A DOI normally contains a prefix beginning with 10. followed by a suffix determined within the registrant's naming system. To a reader, strings such as these have a recognizable appearance.

A language model can learn that appearance.

What it cannot guarantee merely by generating the string is that the identifier was actually registered or that it identifies the publication being discussed.

Those categories can look identical on the page until the identifier is checked.

A Fake DOI Can Accompany a Completely Fabricated Paper

The most straightforward scenario is a nonexistent paper accompanied by a nonexistent DOI.

This fits within the broader problem of AI-hallucinated academic citations. A model can generate plausible authors, an article title, a journal, publication details, and then a DOI-shaped identifier that completes the appearance of a conventional scholarly reference.

None of those elements independently proves that the work exists.

Empirical studies of AI-generated references have documented substantial errors in identifiers. In one evaluation of 500 generated medical references across several AI chatbots, Aljamaan and colleagues classified 227 DOI outputs, or 45.4%, as hallucinated or erroneous under their study conditions. Performance varied markedly among the tools evaluated, so this figure should not be treated as a general DOI error rate for current generative AI systems.

A Real Paper Can Be Paired With the Wrong DOI

DOI hallucination does not necessarily mean the publication is fabricated.

Suppose an AI correctly identifies a real paper, including its authors and title, but supplies the wrong DOI. The citation now contains genuine scholarly information plus one false identifier.

This can happen through a malformed identifier or through confusion with another publication. If you copy the DOI without checking it, readers may be sent somewhere entirely different or encounter an unresolved identifier.

A DOI That Resolves Can Still Be Wrong

This is the crucial verification step.

Crossref's documentation recommends testing a registered DOI by entering its DOI link in a browser and checking whether it resolves correctly. An unregistered DOI should normally lead to a DOI NOT FOUND response.

But successful resolution answers only one question: does this DOI exist?

You still need to ask: does it identify the paper I was given?

If the DOI resolves to a different article, it is still wrong for your citation.

AI Can Generate an Identifier That Looks Highly Plausible

Researchers sometimes expect a fabricated DOI to look obviously nonsensical. That is not a safe assumption.

DOI prefixes and suffixes follow patterns that language models can encounter repeatedly in scholarly text. A model may therefore generate an identifier that resembles those issued by a particular publisher or used by a particular journal.

The result can look especially convincing when the surrounding reference is plausible. You may recognize the publisher, see a familiar DOI prefix, and infer that the identifier is probably genuine.

Pattern conformity is not registration.

Not Every Legitimate Research Paper Has a DOI

A different mistake occurs when researchers expect every legitimate scholarly source to have a DOI.

DOI adoption varies across publication types, publishers, disciplines, and periods. Older literature may lack DOIs. Books, chapters, reports, conference materials, theses, and other legitimate scholarly outputs may or may not have them.

Therefore, the absence of a DOI does not establish that a source is fake.

More importantly, asking a generative AI system to “give me the DOI” for a publication that has none can invite a plausible-looking completion. If an authoritative publication record does not provide a DOI, do not pressure the model to manufacture one.

DOI Verification Does Not Verify the Scientific Claim

Suppose the DOI is correct. It resolves to the right paper. The title and authors match perfectly.

You have verified the identifier and publication. You have not yet verified whatever claim the AI attributed to that paper.

The model can still misrepresent a genuine research paper. It might assign the wrong finding, population, method, statistic, or conclusion to a correctly identified source.

DOI verification is therefore bibliographic verification, not substantive verification.

How to Verify an AI-Generated DOI

The key lesson is simple: a DOI can pass the existence test while failing the identity test.

Common Mistakes When Checking AI-Generated DOIs

Verify the Identifier and the Publication Separately

A DOI is extremely useful precisely because it gives you something concrete to check. Use that advantage rather than treating the identifier as decoration at the end of an AI-generated citation.

Before Using an AI-Generated DOI

Frequently Asked Questions About AI-Generated DOIs

A DOI Is Useful Because You Can Check It

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.