My professor says my sources don't exist. What now?
Before you write a single word of reply, check the references yourself. Open your submitted file, take every reference in the list, and spend the next hour finding out which ones are real. You cannot answer an accusation you have not verified, and the answer changes completely depending on what you find. Send a one-line acknowledgement now ("Thank you, I'm reviewing my reference list and will reply properly by Thursday"), then start the audit.
The panic is the enemy here. The situation is usually smaller than it feels at 11pm, and it is almost always survivable. What sinks people is not the broken reference. It is the reply they send before they know what happened.
The next hour
Freeze the evidence
Save an untouched copy of the exact file you submitted. Do not edit it. Then gather everything that shows how it was written: earlier drafts, your notes, PDFs you downloaded, browser history, and if you used a chatbot, the conversation log. Version history in Google Docs or Word is worth exporting or screenshotting now rather than later. None of this is an admission of anything. It is just the record, and records get lost.
Audit the list, one entry at a time
Run each reference through the standard citation verification checks: does the paper exist, do the details match the publisher's record, and does it say what you cited it for. That guide is the mechanics. What matters here is doing it systematically rather than skimming, and writing the result next to every entry, including the ones that pass.
Start by getting the mechanical pass out of the way. Paste the whole list into our free citation checker, which looks every entry up in CrossRef and OpenAlex and reports two things separately for each one: whether the identifier resolves, and whether the record it resolves to describes the work you cited. Fifty entries at a time, no account, and it takes about a minute. Screenshot the result before you touch anything, because that screenshot is dated evidence that you checked. Then work through the same list by hand for the entries it flags, using the full verification routine. Budget about two minutes each for those.
Sort what you find into three piles
This is the step that turns panic into a plan, because these three failures are not the same and your instructor will not treat them the same.
Pile A: it does not exist. No search result, no DOI, no trace on the publisher's site, no such paper by those authors. This is a fabricated reference, and it is the serious one.
Pile B: it exists, but the details are wrong. Real paper, real authors, wrong year or wrong journal or a mangled title. Often a real paper stapled to somebody else's title. Annoying, fixable, and much less damning than pile A.
Pile C: it exists and the details are right, but it does not say what you cited it for. This is the quietest failure and the most common one. It has its own guide, because "the citation is real" and "the citation is correct" are different claims.
Count each pile. The numbers matter: an accusation that felt like "your whole bibliography is invented" often turns out to be one entry in pile A and four in pile B.
What a failed lookup actually means
Sorting into piles only works if you can read what a lookup is telling you, and a lookup answers two questions that people collapse into one. Does this identifier resolve to a registered record, and does that record describe the work in front of you? Keep them apart and there are five outcomes, each meaning something different.
| What the lookup returns | What it means | Which pile |
|---|---|---|
| The DOI resolves and the record's authors, title and year match | The reference is sound | None. Write "verified" beside it |
| The DOI resolves, but to a different paper | The identifier was transplanted from another real work | A, usually. This is the strongest fabrication signature there is |
| The DOI fails, but a title search finds the work | The paper is real and the identifier was invented or mangled | B. Correctable in one line |
| The DOI fails and no title search finds anything | Neither index holds a work by that title | A, once you have ruled out the row below |
| No DOI in the reference, and no title match | Often a book, chapter, thesis or older article the indexes simply do not carry | None yet. Check this one by hand |
If you would rather see the raw answer than a tool's summary of it, the identifier check is one URL. Put https://api.crossref.org/works/ in front of the DOI in your browser's address bar. A registered DOI returns a JSON record you can read the title and authors straight out of. An unregistered one returns an HTTP 404 whose entire body is the sentence Resource not found., which is about as unambiguous as this ever gets. Checked 11 August 2026.
It helps to know the size of the haystack you are searching. On 11 August 2026 CrossRef's own API reported 185,346,103 records and OpenAlex reported 323,998,640 works. Against numbers like those, a journal article from the last thirty years that appears in neither index is genuinely unusual. A book, an edited-volume chapter or a thesis that appears in neither is completely normal. Which of those two sentences applies to a given entry is the single most useful thing your audit can establish.
That last row is the one that wrongly condemns honest references, and it is the row a panicking student skips. CrossRef and OpenAlex are indexes of works that carry registered identifiers, which means journal articles above all. Monographs, chapters in edited volumes, theses, proceedings from smaller conferences, government and NGO reports, and a great deal of older and non-English scholarship sit outside them. "The index could not find it" and "it does not exist" are different sentences, and writing the second when you mean the first is how a student turns a defensible bibliography into a confession.
A year that is off by one is not a discrepancy. Online-first publication and print publication routinely fall in different calendar years, so the same article genuinely carries two years and different reference managers pick different ones. Our own checker allows a full year of slack before it says a word. Two years apart is worth a look. Four years apart is a different paper.
A matching title is not a matching paper. Titles collide, especially short ones and technical-report titles. Before you accept a title match as a rescue, check that at least one author surname from your reference appears in the record you found. We built that rule into our own matcher as a hard gate: where a reference names two or more authors and the candidate record shares none of their surnames, the match is rejected outright, because a title twin by different authors silently rewrites your citation into someone else's paper and every surface check then passes.
A DOI can fail for reasons that have nothing to do with the paper. A typo in transcription, a character eaten by a PDF copy, an identifier that was registered but whose landing page has since rotted during a platform migration. Five different causes look identical from the outside, which is why the table above separates the identifier's fate from the work's. Before you file anything under pile A, retype the DOI by hand from the original and search the title as well.
The second row deserves its own paragraph, because it looks innocuous and is not. A DOI that resolves cleanly to a real but different paper is very hard to produce by typing and very easy to produce by prediction: a model reaching for a plausible identifier reaches for one it has seen. If your audit turns up an entry in that state, do not describe it as a typo. Describe it accurately and say where it came from.
Our own view, from building the checker that produces that table: the audit is the whole of your defence, and the reason is arithmetic rather than rhetoric. The accusation arrives as a claim about your bibliography, and a bibliography is a set. Nobody who writes "your sources do not exist" has opened all fifteen. They opened two, found them broken, and generalised, which is what a tired marker at 9pm will always do. The only thing that shrinks a claim about a set back to a claim about its members is a member-by-member count produced by you, unprompted, before anyone asks. That is why the mechanical pass comes first and gets screenshotted. A defence that arrives as an argument invites an argument. A defence that arrives as fifteen rows invites a correction.
Knowing the mechanism helps you explain it
If a chatbot was involved anywhere in your process, that is the likely explanation, and it is worth being able to describe accurately: a language model produces references by predicting plausible text rather than by looking anything up, so a realistic author, a believable title and a correctly formatted identifier can all point at nothing. Why AI makes up citations covers the mechanism and how common the failure has become, including in published research by professionals.
Use that to explain, never to excuse. "The AI gave it to me" is not a defence, and framing it as one lands badly. The citation went out under your name, and what you are doing in the reply is showing that you understand that.
How to write the reply
Short, factual, no drama. Four things belong in it:
- What you found. Give the numbers. "I checked all fifteen references. Twelve resolve correctly. Two have incorrect volume and page numbers. One does not appear to exist."
- How it happened. Be accurate and brief. If you used a chatbot to find sources and did not verify them, say so plainly. If you copied a reference from another paper's bibliography without opening it, say that. Inventing a better story is the one thing that reliably makes this worse.
- What you are proposing. A corrected reference list, a resubmission if that is allowed, a meeting. Offer, do not assume.
- The audit itself. Attach it. A table with the reference, whether the DOI resolves, and the corrected details reads as someone taking responsibility.
Do not quietly fix the file and resubmit as though nothing happened. In the best-documented fabricated-citation case in law, two New York lawyers were sanctioned in 2023 less for using ChatGPT than for what they did afterwards: the judge's ruling turned on their attempts to stand behind the fake material once questioned. Editing the record after an accusation is the move that converts a mistake into misconduct.
What happens next
The path varies by institution, but the shape is usually the same: your instructor raises it, there is a conversation or a formal meeting, and someone decides whether a violation occurred. Most universities publish the procedure. Marquette's academic misconduct hearing overview and Indiana University's student code procedures are representative: written notice of the allegation, the right to see the evidence, the right to bring an adviser, a chance to respond, and an appeal route. Find your own institution's version and read it before the meeting, not after.
One thing worth understanding in advance: many integrity codes define fabrication broadly. Iowa State's integrity tutorial and Virginia Tech's definitions of academic misconduct both cover inventing a source and citing a source for a proposition it does not support. Neither definition hinges on you having meant to deceive. Intent tends to shape the outcome rather than the finding, which is exactly why an honest, specific account of how the reference got into your draft is worth more than a defence.
If the accusation is really about AI use rather than the references themselves, that is a different conversation with different evidence, and we cover it in what to do when you're accused of using AI on an essay.
Making sure there is no next time
The habit that prevents all of this is boring and it works: never let a reference into your draft that you have not opened. Not "found in a search result." Opened.
In practice that means verifying as you write instead of at the end. When you add a citation, resolve the DOI in the same minute, and paste the exact supporting sentence into a comment or a footnote alongside it. That quote is your proof later, and writing it down forces you to confirm the source actually says what you think it says. If you are pulling sources from a chatbot, making it search rather than recall helps, but it does not remove the verification step. And when you are starting from scratch, searching real databases beats asking a model to remember papers.
One more thing that pays off later: keep your drafts. A document with a visible history of research, notes and revisions is the single most useful artefact you can have if your work is ever questioned, for any reason.