Ask a modern model what emotion a sentence carries and it will answer instantly, fluently, often correctly. Then watch it use that answer, and you'll see the gap.
It can label the feeling. It cannot feel the weight of it. And in communication, the weight is the whole thing.
Recognition vs. understanding
A model that tags a message "frustrated, 0.92" has performed recognition. Understanding would be knowing that this frustration, from this person, after that week, means say less, not more. Emotion isn't a category. It's context plus history plus stakes, exactly the layer that doesn't survive tokenization.
I'm not a pessimist about this
I build with these tools every day and they're extraordinary. But I think we mislabel what they do. They are spectacular at the surface of language and naive about its purpose. Knowing that keeps me honest about where a human still has to stand in the loop, and, weirdly, more optimistic, because it tells me exactly what's still ours to build.