Models don’t “understand” - they _encode_ the relationships between words.