Google has outlined a three-part method for auditing large language model judges used to evaluate AI outputs. The approach is designed to address bias and inconsistency in automated evaluation.
The guidance argues that large language models used as evaluators are not inherently objective. They can show several forms of bias, including favouring longer answers, reacting differently when options are reordered, and giving undue weight to grammar or formatting. Such systems can also miss incorrect or unsafe outputs and may score responses from related models more highly.
The issue matters because many organisations now use so-called LLM-as-a-judge systems to assess the quality of AI-generated answers, code and other outputs, either in place of or alongside human reviewers. Google argues that the assumption that AI evaluation is more objective than human judgement does not hold up in practice.
Three pillars
The first element of the framework is measuring what Google calls internal stability: whether the same model gives consistent scores when shown the same output more than once.
Because language models are probabilistic, scores can vary across repeated runs even when the text being judged does not change. Teams can test this through multi-pass consistency checks and by calculating score variance, with large swings suggesting the scoring rubric is too ambiguous.
The second element is validation against human judgement. Overall agreement between an AI judge and human reviewers can be misleading if most items in a test set are already valid, because a lenient model may appear accurate while still failing to identify errors.
To address that, Google recommends using chance-corrected metrics to assess how closely an AI evaluator matches expert human decisions. These measures help separate genuine alignment from agreement driven by leniency or random chance.
Model ensembles
The third part of the framework focuses on using multiple judges rather than relying on a single model. One method is decomposed prompting, in which separate evaluators each score a single criterion, such as factual accuracy or tone, instead of one model assessing everything at once.
This structure is intended to reduce the risk that superficial qualities, such as clear formatting or polished writing, influence scores for unrelated criteria. Google links that problem to what it describes as a halo effect in model-based evaluation.
It also recommends ensembling, which means running at least two different models on the same output and comparing their assessments. According to Google, statistical consensus between the models can help teams identify divergence in scoring and spot deterioration in evaluation quality without immediate human review.
The broader argument is that businesses building AI systems should scrutinise evaluators with the same care they apply to the models being tested. Treating automated judges as neutral arbiters without checking their own performance can undermine quality control.
Google backs its position by pointing to published research on common biases in LLM evaluation. These include verbosity bias, where longer and more confident-sounding responses are rewarded even when they are inaccurate, and position bias, where reversing the order of two options can alter the verdict.
It also highlights findings that evaluators may fail to catch faulty or unsafe outputs at a high rate and may overrate outputs generated by themselves or by models from the same family. Those tendencies, Google says, show that model-based assessment should be treated as a subjective process rather than a source of ground truth.
The recommendations come as companies put more weight on automated testing and evaluation in generative AI projects, especially where large volumes of outputs make manual review costly or impractical. For those companies, the challenge is not only building models that perform well, but also making sure the systems used to judge that performance are reliable.
Google said: "Ultimately, LLM judges are not objective arbiters of truth. They are subjective models with their own distinct biases."