AI models get confused by how a table is arranged, not just what it says

It turns out, AI models struggle to understand tables if the layout changes, even slightly. This means systems that use AI to find information in tables can easily miss things just because of how the data is presented.

For years, people assumed that if an AI could read a table, it understood the information regardless of how it was formatted. This paper shows that AI models are surprisingly brittle; they confuse presentation with meaning. This makes it harder to build reliable AI systems that can pull facts from documents, especially when those documents come from many different sources with varying layouts.