And to be clear, this is not a solvable problem. It's not a lack of features. Machine learning models are bafflingly good at mimicking humans, no thing on this earth living or otherwise can do it so well
But to get there they used a trick. They harvested untold articles, books, and conversations, and then created something that given an input would spit out a statistically plausible output from its training set
That approach has a hard ceiling, and you can't bolt the missing bits onto it. It's like how we faked 3D by combining triangles and textures, and with our trickery we got really close to making things look the way they do in life. But just because we got 80% of the way there doesn't mean we're months or years away from completely true to life molecular light pathway rendering
We're stuck with triangles and textures, just like machine learning models are stuck with statistically plausible answers without an iota of actual comprehension