AI fabricates academic citations at scale. Not occasionally, not as a quirk, but as a structural feature of how large language models process text. In January 2026, GPTZero analysed every paper accepted to NeurIPS 2025 and found over 100 hallucinated citations across 51 papers. These papers beat out 15,000 other submissions to get accepted to one of the most prestigious AI conferences in the world. The fabricated citations survived peer review. A Nature investigation asked whether researchers can stop AI making up citations at all. The answer, so far, is no.
The citations AI invents are not random. They use real journal names, real DOI prefixes, plausible author combinations, and titles that sound so inevitable you assume someone must have written the paper. The conventions of academic publishing are the disguise.
In this post we will:
Give you a game that tests whether you can spot fabricated citations.
Show what happened when a reader searched for a fake citation and got something worse than a dead link.
Argue that verification is a skill, not an instinct, and that most of us have never been taught it.
How to spot fake AI citations (and why instinct is not enough)
I built Dead Reference to make this problem tangible. It is a free browser game. You are shown 10 citations drawn from a pool of 40: half real, half fabricated. Your job is to identify which are which.
The real citations include papers from Nature, The Lancet, and the ACM Conference on Fairness, Accountability, and Transparency. The fake ones include papers in the Review of Educational Research, the Journal of Media Psychology, and The Lancet Digital Health. Every fake citation uses the correct DOI prefix for its publisher. Every fake title could plausibly exist.
Most players score around 5 or 6 out of 10 on their first attempt. That is barely better than guessing.
The game is not trying to embarrass you. It is trying to show you that academic citation is a system built on trust, and that AI has learned to exploit that trust with precision. You cannot spot a fabricated citation by reading it carefully. You can only spot it by checking it against an external source: a DOI resolver, Google Scholar, a publisher catalogue. Instinct is not verification. The game teaches the difference.
What happens when you search for a citation that does not exist
Last week, Kristina Kroot played Dead Reference and then did something I had not anticipated. She searched for one of the fabricated citations using a search engine with AI-powered summaries.
The search engine did not return a dead link. It did not say the citation was not found. It generated a confident summary of the nonexistent book, complete with themed arguments and supporting citations. The fabricated reference was treated as real, summarised as real, and presented to Kristina as if it had always existed.
She had to scroll past the AI summary to find the absence. As she put it: “Most people won’t scroll.”
This is the second layer of the problem. AI does not just fabricate citations. AI search engines then validate those fabrications. The loop closes. A student generates a citation using ChatGPT. The citation does not exist. The student searches for it. The search engine summarises the nonexistent source as if it were real. The student includes it in their essay. The citation enters the academic record. Other AI models train on it.
This is not a hypothetical. This is happening now, in every discipline, at every level of education.
What a fabricated AI citation looks like in practice
I built Dead Reference because I wanted to turn a research finding into a felt experience. The data says AI fabricates citations. Reading that sentence does not change behaviour. Playing a game where you confidently mark a fake citation as real, and then discover you were wrong, does.
One of the fabricated citations in the game reads:
Zhang, W., & Patel, S. (2021). Algorithmic bias in automated essay scoring: Implications for equity in higher education. Review of Educational Research, 91(4), 512-538.
I work in higher education. I study AI in education. This title is so plausible that I would not question it in a literature review. The Review of Educational Research is a real, top-tier journal. The DOI prefix is correct for SAGE/AERA publications. The topic is timely. Everything about it feels inevitable.
The paper does not exist. The authors do not exist. The DOI does not resolve.
That is the game teaching you what the research cannot: the feeling of being fooled by something you should have checked.
What to share
Play Dead Reference this week. When you finish your first round, notice which citations you got wrong and why. Did the journal name convince you? The author’s initials? The year? Share your score in the comments and tell us which citation fooled you. If you searched for a fake citation and found an AI-generated summary confirming it, we want to hear about that too.
Rebecca Wicker, author of The Strategic Linguist, published a piece this week that puts the fabricated citation problem in a broader frame. Her argument: the mechanism that turned a single 1960s psychology study into a universal law of communication (“93% of communication is nonverbal”) is the same mechanism now operating across AI. Findings get extracted from context, stripped of caveats, and repeated until the repetition becomes the evidence. Rebecca calls fabricated AI citations “the Mehrabian problem on an industrial scale: the performance of evidence, indistinguishable on the surface from evidence itself.”
Rebecca also featured Dead Reference in the piece and described the game better than I have:
Why checking AI citations is a critical literacy skill
The infrastructure of academic trust was built for a world where fabricating a citation required effort. You had to invent an author, a journal, a volume number, and hope nobody checked. AI removed the effort. It generates citations that are structurally perfect and referentially empty. The conventions that once signalled credibility now signal nothing.
This is not a problem that detection tools will solve. It is a problem that verification habits will solve. The skill is not recognising a fake citation by looking at it. The skill is refusing to trust any citation until you have checked it against a source that exists outside the AI’s training data.
That is what critical AI literacy means in practice. Not knowing how to use AI. Knowing when to stop trusting it.
If you play Dead Reference, I would love to hear your score and which citation caught you out.
I read and respond to all your comments.
Go slow.
This post is part of Slow AI’s ongoing work to help you engage with AI critically, not just learn to use it faster.
This month in the curriculum: 166 paid members explored how AI image generators default to colonial tropes when depicting race, profession, and power. We ran the same prompt across tools and continents. The biases were near-identical. The conversation that followed was not.
The Slow AI Curriculum for Critical Literacy is a 12-month programme for people who want to learn when AI helps, when it harms, and how to tell the difference. Each month covers one critical theme, with a live 45-minute seminar, research syntheses, moderated dialogue, and full recordings. You receive a Certificate of Critical AI Literacy upon completion.
166 members from education, research, policy, and industry.
Read the 2026 Handbook for Critical AI Literacy to see exactly what the programme covers. £75/year.
i love all the gamification Sam. excellent stuff here!
Not only do you help us understand what's happening but you give us a free tool to use to help us practice this skill!? 🤯 Thank you! Dead Reference is such an incredible tool. No matter your score, it's really eye opening to see how we take things at face value because we have a preconceived idea of what it 'should' look like.