When AI Can't Read, It Invents — But It Still Sees the Shape
Latest AI models struggle with low-resolution document reading, risking accounting errors with fabricated details.
Recent tests revealed that the latest AI models in low-resolution modes failed to accurately extract information from a synthetic invoice. While they returned correct figures, critical details like the customer name were fabricated. This raises concerns about potential errors in accounting processes, where clean arithmetic could lead to unchecked payments based on incorrect identities.