Using AI models to grade other AI outputs is now standard practice — except it may be quietly failing for most of the world's languages.
A new paper reviewed 650 research papers that used large language models as evaluators of natural language generation. Only 33 of those papers examined low-resource or multilingual settings. Among those 33, the researchers found inconsistent results, a near-universal habit of relying on a single judge model, and a tendency to accept AI verdicts without adequate human verification — the same human verification that would reveal whether those verdicts were any good.
The problem compounds itself: LLMs already struggle with low-resource languages, meaning the models being evaluated are weaker in those languages, and the models doing the evaluating are weaker too. Stacking one unreliable system on top of another and calling it a methodology is a risk the NLP field appears to be taking with minimal scrutiny. The paper's authors offer a set of recommendations, though the core fix — more human validation — is also the expensive one nobody wants to pay for.
This is worth watching because LLM-as-a-Judge has become the path of least resistance for evaluation at scale. If the approach is systematically biased toward high-resource languages like English, then benchmarks built on it are measuring how well AI performs for a slice of the global population and presenting the results as universal. That's not a niche methodological footnote — it's a quiet distortion built into how the field measures progress.