MWS AI, a division of MTS Web Services, has evaluated the ability of multimodal artificial intelligence models to detect fraudulent documents. The assessment utilized the MWS AI Vision Bench, an open benchmark designed to measure performance in optical character recognition and document comprehension.
The research team expanded the benchmark with an Anti-fraud dataset containing 430 images. This collection includes 200 original documents, 130 manually edited files, and 100 images generated by AI. Models were tasked with classifying the images and identifying suspicious elements, such as altered details or AI-generated content.
The results indicate that models proficient in general document processing do not always excel at fraud detection. For instance, while GPT-6.1 Sol leads in overall performance, it ranks third in fraud detection, where Claude Opus 5.5 currently holds the top position. Other top-performing families include GLM.
According to benchmark creator Georgy Gaykov, the Anti-fraud task is experimental and currently excluded from the overall MWS AI Vision Bench score. He noted that while the dataset helps compare universal models, specialized solutions remain more effective for real-world fraud protection.
