When an AI model grades other AI answers, how much of the 'score changed on rerun' problem is the grader's fault, not the answer's?

This explores how much of the wobble in prompt evaluation (scores that change when you rerun or rephrase things) comes from the evaluator, as opposed to the thing being evaluated. The corpus suggests the judge accounts for a large share, and possibly most of it on complex tasks. One striking number: on hard tasks, standard LLM-as-a-Judge setups showed a 31% 'judge shift,' meaning their verdicts changed across runs. An agentic evaluator that actively gathers evidence before ruling cut that to 0.27% Can agents evaluate AI outputs more reliably than language models?. A hundredfold drop from changing only the evaluator is strong evidence that much of the instability lived in the judge, not in the outputs being judged.

The biases themselves are fairly well mapped. They include position (favoring whichever answer comes first), verbosity (longer looks better), authority cues, and surface 'beauty.' Human evaluators share a version of this. In one study, people rated polished AI-written documents as better than human ones, and also mistook them for human work Does polished writing actually signal better quality work?. The fix most people try first, telling the judge to be fair, doesn't reliably work Can prompting reduce bias in LLM judges reliably?. What does seem to work is changing how the judge reasons. Judges trained with reinforcement learning to think through a verdict, rather than react to surface features, become much less susceptible to these biases Can reasoning during evaluation reduce judgment bias in LLM judges?. Breaking an output into fragments and checking each one against the criteria also helps: practitioners caught 48% more misalignments than a single overall score revealed Can breaking outputs into fragments reveal more evaluation problems?. Holistic scores hide disagreement that is really there.

The judge is still not the whole story. Some instability is real and lives in the model being tested. Models with low confidence on a task swing widely when a prompt is reworded, while confident models hold steady. Larger models, few-shot examples, and objective tasks all raise that confidence Does model confidence predict robustness to prompt changes?. Which prompt technique 'wins' also depends on the model tier: step-by-step reasoning helps cheap models but can hurt strong ones Do prompt techniques work the same across all LLM tiers?. So a prompt that looks unstable may be accurately reflecting an uncertain model, not a biased judge. One essay goes further and argues that this variability is a built-in property of AI outputs, not a defect that better quality assurance can remove Why does AI output change with every prompt and context?.

There is a third source people often miss: the human who iterates on the prompt. When one researcher keeps revising a prompt until results look right, the evaluation criteria quietly drift toward what the model can already do, and the loop confirms itself Does iterative prompt engineering undermine scientific validity?. That is evaluator bias too, just upstream of the LLM judge. One remedy is to fix the criteria in advance. Another is to score prompts on their own merits, using structured dimensions such as clarity, logic, and instruction quality, rather than only on what they produce Can we measure prompt quality independent of model outputs?.

Here is the takeaway you may not have expected. The corpus doesn't give a single percentage, but it shows the instability has layers. There is judge bias (large and reducible through structure, not instructions), model uncertainty (real signal worth measuring), and researcher drift (invisible unless you pre-register criteria). If your prompt evaluations are unstable, first test whether a different evaluator setup stabilizes them. If it does, the problem was the judge. If it doesn't, the instability is telling you something about the model.