AI judges keep picking machine-written arguments over human ones, even when people don't — so where does that bias come from?
Why do LLMs systematically prefer text from their own family?
This explores why LLM judges tend to pick machine-written text over human-written text, and what the corpus can and can't say about the cause.
This explores why LLM judges tend to pick machine-written text over human writing. The corpus documents the effect well but only hints at the cause. When judging debates, LLM judges chose the LLM's argument as the winner 62% of the time, against 37% for the human's. Human readers split almost evenly, 39% for the LLM and 37% for the human Do LLM judges systematically favor arguments from other LLMs?. The tilt comes from the judge, not from the text. One limit on the question as asked: that study compares LLMs against humans. It doesn't test whether a model favors its own family over rival models, so the corpus supports 'LLMs prefer LLM text' more firmly than 'LLMs prefer their own family's text'.
The preference also doesn't seem to track quality. A meta-analysis of 7 studies with over 17,000 participants found LLMs and humans about equally persuasive (Hedges' g = 0.02) Are language models actually more persuasive than humans?. If machine arguments were simply better, human readers would notice too. The judge is responding to something people mostly aren't.
The likeliest candidate is surface features. LLM judges fall for fake references and rich formatting, and those biases work regardless of what the text says Can LLM judges be fooled by fake credentials and formatting?. LLM writing also has a recognizable fingerprint. Models recover only about half the distinct arguments humans make and keep reusing the same hedged sub-arguments Do language models flatten the range of public arguments?. They use about 22% more moral language than humans while matching human sentiment scores Do LLMs use moral language more than humans?. Each register inherits its habits from its training distribution Why do LLMs produce such different writing in chat versus posts?. A judge trained on similar data may find that fingerprint comfortable, but that link is an inference. No note in the corpus tests it.
The bias may also be baked in early. LLM recommenders show position, popularity and fairness biases that come from the pretraining objective and corpus, not from any task-specific data Where do recommendation biases come from in language models?. If self-preference works the same way, fine-tuning a judge won't remove it. That matters for AI-feedback training pipelines, where the judge's favoritism sets a ceiling on how well the feedback can be calibrated Do LLM judges systematically favor arguments from other LLMs?.
The corpus has no direct test of same-family versus cross-family preference, and none that measures 'familiarity' inside the judge. The question of why is still open here.
Sources 7 notes
LLM judges selected LLM arguments as winners 62% of the time versus humans' 37%, while humans split votes 39% LLM / 37% human. This same-author bias operates downstream of component scoring and compounds existing judge vulnerabilities, creating a calibration ceiling in RLAIF pipelines.
A meta-analysis of 7 studies with 17,422 participants found no detectable difference in persuasive effectiveness between LLMs and humans (Hedges' g = 0.02). Persuasiveness appears conditional on context rather than speaker category.
Research identified four evaluation biases in LLM judges, with authority and beauty biases being semantics-agnostic and trivially exploitable through fake references and formatting—zero-shot attacks requiring no model access or optimization.
Across 23,384 LLM essays on debates, models recover only half of distinct human arguments and reuse hedged sub-arguments—a gap in argumentative structure, not just prose style. Diversity prompting adds noise outside human argument space rather than filling the long tail.
Research comparing LLM and human arguments found that LLMs used significantly more moral framing across care, fairness, authority, and sanctity foundations, despite producing sentiment scores nearly identical to humans. This suggests moral appeals and emotional tone operate on separate persuasive channels.
Show all 7 sources
The same model produces sycophantic chat (shaped by RLHF on conversational data) and falsely objective posts (shaped by published prose training). Each register inherits failure modes from its training distribution rather than representing different models or subsystems.
Wu et al. show that LLM-based recommendation systems exhibit position bias, popularity bias, and fairness bias—unique failure modes stemming from the language model's pretraining objective and corpus demographics rather than interaction data. Mitigation requires LLM-specific approaches, not adapted collaborative filtering techniques.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- Large Language Models are as persuasive as humans, but how? About the cognitive effort and moral-emotional language of LLM arguments
- Do LLMs Change Their Minds Like Humans? Diagnosing Human--LLM Divergence in Single-Turn Persuasion Judgments
- Argument Collapse: LLMs Flatten Long-Form Public Debate
- The Thin Line Between Comprehension and Persuasion in LLMs
- A meta-analysis of the persuasive power of large language models
- Humans or LLMs as the Judge? A Study on Judgement Biases
- Can Language Models Recognize Convincing Arguments?
- When Large Language Models are More Persuasive Than Incentivized Humans, and Why