A new benchmark called WildHandBench has been created to evaluate the ability of artificial intelligence models and humans to understand handwritten text. The benchmark consists of 500 handwritten documents in three structures (free text, tables, formulas), four languages, and nine real-world scenarios. It also introduces a metric called Prior-Driven Error (PDE) that measures whether errors are due to language priors or visual evidence. Evaluations using 18 state-of-the-art models and human baselines show that the best model achieves only 71.85% accuracy, while humans outperform all models with 77.09% accuracy. The results also indicate that AI models tend to rely heavily on language priors, which conventional metrics may not capture. --- Why it matters: This matters because it highlights the limitations of current AI models in understanding handwritten text and their reliance on language priors, which can lead to systematic errors. It also provides a new benchmark for researchers to evaluate and improve these models. Source: https://arxiv.org/abs/2608.22959