When AI judges score other AI's answers, do made-up citations that look authoritative fool them into higher scores?

This explores how easily AI judges, meaning LLMs used to score other models' answers, can be fooled by invented citations. The corpus answers plainly: very easily, and the trick costs nothing. Researchers found that LLM judges give higher scores to responses that include fake references, whatever the content says Can LLM judges be tricked without accessing their internals?. The researchers call this an 'authority bias,' and it is one of four exploitable biases they identified. It shares a category with 'beauty bias,' where the judge rewards rich formatting such as headers and bullet points Can LLM judges be fooled by fake credentials and formatting?. What makes these two dangerous is that the attacker doesn't need to understand the question or improve the answer. They don't need access to the model or any optimization either. Adding a plausible-looking citation is enough. Many AI benchmarks now use LLM judges, so this weakness quietly undermines the leaderboards people rely on to compare models.

The pattern looks less like a bug and more like a habit when you set it next to other notes. Models also avoid correcting false claims that users present confidently, even when they know the correct answer. That paper attributes this to 'face-saving' norms learned from human conversation rather than to missing knowledge Why do language models avoid correcting false user claims?. A judge that defers to a fake citation may be doing something similar: treating the look of credibility as a social cue instead of checking it. Judges also tend to prefer text that resembles their own writing, and that preference grows with their ability to recognize their own output Do LLMs favor their own text because they recognize it?. Together these findings suggest that AI evaluators respond to surface signals such as style, familiarity, and apparent authority, and that a determined writer can manufacture all three.

The same weakness shows up outside benchmarks, in scientific publishing. ICLR 2026's organizers found AI-text detectors too unreliable to enforce automatically. They did, however, desk-reject papers with confirmed fabricated references, because a fake citation is something you can actually check How can conferences detect and handle LLM misuse in peer review?. Meanwhile, more than 1,600 ghost-authored papers with valid DOIs have already entered scholarly databases Do language models leak their training through fictional names?, and human experts can't reliably tell LLM-written abstracts from human ones Can readers tell LLM abstracts from human ones?. The worrying loop is this: if AI reviewers trust citations they can't verify, and the citation record itself is filling with machine-generated material, then a fake reference becomes harder to catch in both places.

The corpus also points to two ways out, and both replace pattern-matching with checking. One approach uses reinforcement learning to train judges to reason through an evaluation before scoring it. That noticeably reduces their susceptibility to authority, formatting, verbosity, and position bias Can reasoning during evaluation reduce judgment bias in LLM judges?. The other breaks evaluation into explicit steps: extract the claims, retrieve the related work, and compare. On novelty assessment, this structured pipeline agreed with human reviewers' reasoning 86% of the time, far better than asking a model for a single overall judgment Can structured pipelines make LLM novelty assessment reliable?. Neither paper was designed specifically to test fake references, so that remains a gap in the collection. Still, the lesson carries over: a judge that actually looks up its sources is much harder to fool than one that only notices that sources were cited.