ChatGPT gave me a DOI that doesn't work. Which kind of broken is it?

Clean the string and try once more. Strip the trailing full stop, the doi: label, any stray space, and the hyphen your PDF reader inserted at a line break, then paste the bare identifier after https://doi.org/. That fixes a surprising share of failures on the spot. If it still fails, you have one of five distinct problems, and only one of them means the paper is fake.

This page is about the DOI itself: what its failure mode tells you, and how to read it. If your real question is whether the article exists at all, that is the general citation verification routine and it takes about a minute. Come back here once you know the answer, because the answer is what turns a dead link into a diagnosis.

Why the format proves nothing

A DOI has two parts. The prefix always begins 10. and identifies the organisation that registered it, so 10.1037 belongs to the American Psychological Association and 10.1038 to Springer Nature. The suffix is whatever the publisher chose, and Crossref advises its members to make suffixes opaque, essentially meaningless strings, so that nobody tries to read information into them. DataCite, the registration agency used for datasets and other research outputs, describes the same shape.

Two consequences follow, and both matter here. A DOI is real only because it was registered, not because it is well-formed: there is no checksum, nothing that makes an invalid DOI look invalid. And because the suffix carries no meaning, a model that has seen millions of DOIs can produce a new one that follows every convention perfectly. Format is free. Registration is not.

The five causes, and how each one looks

1. The string got damaged in transit

What you see: "DOI Not Found," and then it works after you clean it.

The commonest cause by a distance, and it has nothing to do with AI. Suspects: a full stop carried over from the end of a sentence, a doi: prefix from an older citation style, a hyphen inserted by line wrapping inside a PDF, a curly quote, a non-breaking space, a suffix truncated where a column ended. Crossref's display guidelines now recommend writing DOIs as full clickable https://doi.org/… links largely to stop this happening.

2. The registration has not propagated yet

What you see: "DOI Not Found," but the article is right there on the journal's website, published within the last few weeks.

There is a short window after publication where a newly deposited DOI has not become resolvable everywhere. If the article's landing page exists and carries a recent "online first" or "published" date, this is almost certainly what you are looking at. The reference is fine. Cite the DOI as printed on the article page and note that the link may take a little while to work.

3. A broken or migrated registration

What you see: "DOI Not Found," and the article exists on the publisher's site with that exact DOI printed on it, sometimes years old.

Publishers move journals between platforms, transfer titles between owners, and occasionally fail to deposit or maintain a registration. Crossref's own support forum carries a steady stream of these reports, including threads on fixing DOI Not Found after a hosting change. Crossref also compiles a daily error report of failed resolutions and sends it to the publisher that owns the prefix, on the assumption that some of them are legitimate DOIs someone forgot to register. Your reference is correct; the infrastructure is not. Cite the DOI and link the publisher's article page as well.

4. It resolves, but to a different paper

What you see: no error at all. A page loads, and it is not your article.

People miss this one because they check that the link works rather than where it goes. Two causes. The citation may quote the DOI of the whole issue, the whole book or the parent record rather than the specific article or chapter, which is a common slip when references are copied by hand. Or an invented suffix has landed on a string that happens to be registered to something else. Either way, trust the page over the citation: the DOI is the authoritative identifier, so if they disagree, the rest of the reference was assembled around the wrong one.

5. It was never a DOI

What you see: "DOI Not Found," no article anywhere under that title, and often two or three other references in the same list failing the same way.

Fabricated DOIs cluster. If one reference from a batch fails this test, check the others rather than assuming you found the only bad one. The reason lies in how the reference was produced: a language model generates a DOI the way it generates any other text, by predicting a plausible sequence, which is the same mechanism behind fabricated references generally.

The check that separates 3 from 5

Cause 3 and cause 5 look identical from the outside: a dead link and a well-formed string. One check pulls them apart in ten seconds, and hardly anyone knows about it.

Open https://api.crossref.org/prefixes/10.1037 in your browser, swapping in the prefix from your DOI. It returns the name of the organisation that owns it. Now compare that against the journal named in your citation. A real DOI's prefix belongs to the publisher of that journal, or to a publisher that once owned it. An invented DOI usually borrows a familiar prefix and bolts an invented suffix onto it, so you get results like a prefix owned by a medical association attached to a citation for an education journal. That mismatch is the strongest single signal available without finding the paper.

The check has a limit worth knowing: journals do change publishers, so an old article can legitimately carry a prefix belonging to a previous owner. Treat a mismatch as strong evidence rather than proof, and confirm it against whether the article exists at all.

The decision tree

1. Clean the string and retry. Works now? Cause 1. Done.
2. Did a page load? If yes but it is the wrong article, cause 4. Rebuild the reference around the paper you meant.
3. Check the prefix owner. Unrelated to the journal named in the citation? Strong signal for cause 5. Keep going to confirm.
4. Does the article exist? Establish this with the standard verification routine. You only need a yes or a no.
5. Yes, published in the last month or so: cause 2, and the reference is sound.
6. Yes, and that DOI appears on the article's own page: cause 3. The registration is broken, not the citation.
7. Yes, but the article's real DOI is different: the one you were handed was wrong. Take the registered one and re-check the other fields, because a model that got the DOI wrong usually got the year or volume wrong too.
8. No article anywhere: cause 5. Delete the reference and check the rest of the list.

After the diagnosis

If the paper is real, you still have two jobs left, and the DOI told you nothing about either. One is whether the reference's other fields match the publisher's record. The other is whether the paper actually supports the sentence you attached it to, which is a separate and much more commonly failed test that we go deep on in when the source exists but says something else. A resolving DOI also says nothing about whether the paper still stands, so check its retraction status while you are on the article page.

If the DOI was invented and the work has already been submitted, the guide you want is what to do when your professor says your sources don't exist.

Why you cannot prompt your way out of this

Asking a chatbot to "only give real DOIs," or to confirm that a DOI it just produced is valid, does not work, because the model has no way to query the registry. It can only generate more text that looks like a confirmation. Making the tool search the web genuinely helps, since then the identifiers come from pages it just read, and we cover how to set that up in how to get ChatGPT to cite real sources. It does not remove the resolve step.

The habit that costs nothing: paste each DOI into doi.org the moment it enters your draft, and glance at the page that opens. One tab, one paste, two seconds. A DOI that has been through that once will never surprise you the night before a deadline.