AI judges tend to pick AI-written answers over human ones, while human judges don't, so why the tilt?
Why do LLM judges systematically favor outputs from their own model family?
This explores why AI judges lean toward text written by AI, ideally by their own model family, and what the collection can and can't say about the cause.
This explores why AI judges lean toward AI-written text, and specifically their own model family's. The closest evidence is a debate-style test. LLM judges picked the LLM-written argument as winner 62% of the time and the human one 37% of the time, while human judges split almost evenly (39% LLM, 37% human) Do LLM judges systematically favor arguments from other LLMs?. Humans didn't share the tilt, so it comes from the judge, not from the LLM arguments being better. The collection has no note comparing a model's favoritism toward its own family with its treatment of a rival's outputs. The finding is LLM versus human, and no note explains the cause.
Neighboring notes suggest a likely mechanism, though this is an inference and not a tested result. LLM judges are easy to sway with surface cues. Fake references and rich formatting raise scores regardless of content quality, and the attack needs no access to the model Can LLM judges be fooled by fake credentials and formatting? Can LLM judges be tricked without accessing their internals?. If a judge rewards how an answer looks and sounds more than what it says, then text with a machine's own polish will score well. The same-author bias shows up at the final pick, after the individual components are scored, and it stacks on top of these other weaknesses Do LLM judges systematically favor arguments from other LLMs?.
The stakes are highest when AI judges train other AIs. In RLAIF pipelines the same-author bias creates a calibration ceiling. A judge that rewards its own style can only steer a model toward what the judge likes, and that isn't necessarily what people would choose Do LLM judges systematically favor arguments from other LLMs?.
Instructing a judge to be unbiased doesn't reliably work Can prompting reduce bias in LLM judges reliably?. Two other approaches have more support. One is training judges with reinforcement learning to reason before scoring. This cut susceptibility to authority, verbosity, position and beauty bias Can reasoning during evaluation reduce judgment bias in LLM judges?, but self-preference isn't among the biases it was tested on. The other is a set of mechanical safeguards that don't depend on the judge's own honesty. These include measuring the judge against human labels and planting known cases as alarms Can deterministic checks protect LLM judges from failure?. The human-label check is the one that would catch own-family favoritism, since humans are the only reference point that doesn't share the bias.
Sources 6 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.
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.
Research shows LLM evaluators systematically score higher when responses include fake references or rich formatting, independent of content quality. These biases are exploitable without model access, undermining AI benchmark credibility.
Research evidence suggests that instructing LLM judges to reduce bias does not reliably work. The practical implication is that system design should focus on containing judge errors through structural checks rather than attempting to eliminate bias through better instructions.
Training judges with reinforcement learning to reason about evaluations—by converting judgment tasks into verifiable problems with synthetic data pairs—produces judges that think through their decisions rather than relying on exploitable surface features, directly mitigating authority, verbosity, position, and beauty bias.
Show all 6 sources
Research identifies four mechanical safeguards: ordering unarguable checks before contestable ones, measuring correctness against human labels, hiding test data from proposers, and using planted cases as alarms. None requires the LLM itself to verify compliance.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- Humans or LLMs as the Judge? A Study on Judgement Biases
- Learning to Plan & Reason for Evaluation with Thinking-LLM-as-a-Judge
- Can You Trust LLM Judgments? Reliability of LLM-as-a-Judge
- ChatEval: Towards Better LLM-based Evaluators through Multi-Agent Debate
- LLM-as-a-Judge Is Not an Oracle: Why Self-Improving Agents Need Deterministic Guardrails
- When Reject Turns into Accept: Quantifying the Vulnerability of LLM-Based Scientific Reviewers to Indirect Prompt Injection
- The Thin Line Between Comprehension and Persuasion in LLMs
- Could you be wrong: Debiasing LLMs using a metacognitive prompt for improving human decision making