The first answer is plain and correct. The second is elegantly arranged, reassuringly confident, and wrong in one important place. You ask another AI which is better. It picks the elegant one and writes a persuasive explanation of its choice.
This is an illustrative failure case, not a result from a test we performed. It captures a problem facing anyone who uses a model to rank draft copy, compare research answers, or decide whether a support response is ready. The evaluator brings preferences of its own.
The 2023 MT-Bench and Chatbot Arena paper explored both the usefulness and limits of language-model judges, including biases related to order, verbosity, and a model's own outputs. The April 2026 first version of a study compared several mitigation approaches across five judges and reported that presentation style was a particularly strong influence in its tests. This article refers specifically to that preprint version, not a settled verdict about every evaluator.
The lesson is practical: before trusting the judge's ranking, find out what it is rewarding.
Give correctness somewhere to stand
Start with a small task where a person can establish the essential facts. Imagine a fictional museum page that says admission is free on Thursdays and costs $12 on other days. One answer preserves both facts. Another says admission is always free and adds a polished itinerary.
A useful evaluator should fail the second answer on the admission claim even if it prefers the tone. Keep the source passage beside the candidate answers and ask the judge to identify the exact claim that violates it. A broad score for “helpfulness” can otherwise trade factual accuracy against charm without making the exchange visible.
Separate the questions. Is the answer factually supported? Does it cover what the user asked? Is it clear? Only then ask which version is preferable overall. If one dimension is essential, say so explicitly. The grading instruction should not leave the model to decide whether a useful falsehood earns enough style points to pass.
Change the wrapper, keep the meaning
Make a second version of each answer with the same factual content but different presentation. Turn the polished answer into a plain paragraph. Give the plain answer headings and smoother transitions. Do not improve or damage the claims while changing the wrapper.
Now ask whether the verdict changes. If it does, inspect the judge's explanation rather than treating the difference as automatically irrational. Presentation can matter to readability. The question is whether it overrode a stated requirement such as factual correctness.
Also reverse the order of the two answers. An order-sensitive choice is a useful warning when the underlying task and content have stayed fixed. It does not tell you which answer is right. It tells you that one comparison did not establish a dependable preference.
These are proposed checks for your own evaluator. They do not reproduce either cited paper and do not estimate a general bias rate.
Keep a few deliberately difficult cases
The most informative examples are often close calls. A response may cite the right document but change “may” to “will.” It may preserve a price while dropping the condition attached to it. It may answer the obvious part of a question while ignoring the exception the reader actually needs.
Create a handful of such cases from public or invented material. Record the expected judgment before running the model. Include ties and cases where neither answer should pass. Otherwise the evaluator can appear decisive simply because it has never been allowed to reject the whole pair.
Compare the judgments with more than one person's reading when the standard is subjective. A disagreement about humor or tone may represent a real editorial choice. A disagreement about a date in a source is a different problem. Keeping those differences visible makes the evaluator easier to use intelligently.
Use the judge to find work for a person
A model can still be valuable here. It can point to a contradiction, flag an unsupported sentence, or explain why two answers may serve different readers. Those are specific contributions a reviewer can examine.
Treat a verdict as a route into the evidence rather than a substitute for it. Ask the evaluator for the decisive passage and the particular rule it applied. If the explanation cannot survive comparison with the source, another round of fluent justification does not make the decision stronger.
For a large collection, use the judge to organize attention: group obvious failures, identify close comparisons, and surface recurring ambiguities in the instructions. Keep a human-reviewed sample and inspect the cases where the model disagrees with it. The disagreements may expose a weak judge, an unclear task, or an editorial standard that needs saying aloud.
An AI judge becomes more useful when it is allowed to reveal its limits. The goal is not a machine that always produces a winner. It is a review process in which the winning answer has a reason that survives being read.
Sources and limitations
- Zheng et al., “Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena”: checked October 15, 2026. Primary research on model judging and documented bias types; results concern the models and tests studied.
- Soumik, “Judging the Judges: A Systematic Evaluation of Bias Mitigation Strategies in LLM-as-a-Judge Pipelines,” version 1: checked October 15, 2026. April 25, 2026 preprint version examining bias and mitigation across specified models and benchmarks; later publication status is not characterized here. Not independently replicated by this publication.