Why LLMs Double Down on Made-Up Citations

1. Citations are a style, not a lookup

An LLM does not retrieve references from a database. It predicts text one token at a time, and academic citations are an extremely regular pattern: Author, Year, plausible title, real journal, volume, pages. The model has seen millions of them, so it can produce a flawless-looking one for any topic, whether or not the paper exists. The demo below fabricates citations the same way: by filling a learned template with statistically plausible parts. Every one is fake, and every one looks real.

2. Why calling it out often makes things worse

Three forces stack up when you say "that citation is fake":

1. Context consistency pressure. The fake citation is now in the conversation history. The model's training strongly rewards text that is coherent with prior context. Continuing to treat the citation as real is, token by token, the more "probable" continuation than a sudden reversal.
2. Confidence was rewarded in training. Human raters preferred fluent, assured answers. Hedging and "I don't know" were often scored lower, so the model learned an assertive register — even when the underlying content is fabricated.
3. No self-model, no ownership. The model does not view the citation as "its work" because it has no persistent sense of authorship or belief. It cannot check a memory of having looked the paper up, because no lookup ever happened. It only sees text that asserts the paper exists, written in a confident voice — so it continues in that voice.

So doubling down is not stubbornness or ego. It is the path of least resistance through probability space: the transcript so far says the citation is real, and confident prose is the house style.

3. The doubling-down simulator

Slide the consistency-pressure control and challenge the model. High pressure (long confident context, assertive style) makes retraction unlikely. Low pressure (short context, hedged style, or an explicit request to verify) makes admission far more likely — which matches real-world prompting experience.

4. What the model actually optimizes

At every step the model picks from a distribution over next tokens. After a confident fake citation appears in context, tokens that continue the confident storyline carry more probability mass than tokens that begin a retraction. This chart is illustrative, not measured — but it is the shape of the mechanism.

5. What actually reduces it

Because the failure is mechanical, the fixes are too: ask the model to verify against retrieved sources rather than memory; request hedged answers up front ("say unsure if unsure"); start a fresh conversation instead of arguing inside the poisoned context; and treat any citation you cannot click through to as unverified. Newer systems that ground answers in retrieval hallucinate references far less, precisely because a lookup finally happens.

Short answer to the original post: no, the model does not view the citation as its own work. It has no view at all. It has a transcript that asserts the paper is real and a training history that rewarded sounding sure — and it keeps writing the most probable next sentence.