Researchers found that AI judges fail to accurately measure workplace performance

A new research paper titled O*NET-BENCH reveals a significant gap between how artificial intelligence evaluates work tasks and how human workers actually rate them. The study tested 33 different model configurations across six model families using 4,501 real-world worker ratings.

While many AI models are good at ranking tasks in the correct order, this ability does not mean they agree with humans on whether a job output is acceptable. The AI models estimated acceptable output rates anywhere from 3% to nearly 98%, whereas human workers consistently rated about 61% of tasks as acceptable.

Imagine you ask a digital assistant to judge if a draft report is professional enough for a client. The AI might correctly rank it as better than a messy draft, but it may have a completely different threshold for what 'good enough' actually means compared to a human employee.

The researchers discovered that even when they calibrated the models to remove bias, the AI scores explained less than 9% of the differences in individual human ratings. Simple statistical methods, like a basic text-matching baseline, performed nearly as well as the advanced AI models at predicting these rankings.

Could we ever rely on AI to objectively measure human productivity in the office if its standards for quality are so disconnected from our own?

The study is limited to the specific O*NET-BENCH dataset derived from 45,796 worker ratings. It does not prove that AI models are incapable of improvement, only that current ranking-based evaluation methods are not enough to predict human-level acceptance in professional settings.