The Matplotlib team are completely in the right to ban AI. The ratio of usefulness to noise makes AI bans the only sane move. Why waste the time they are donating to a project on filtering out low quality slop?
They also lost nothing of value. The 'improvement' doesn't even yield the claimed benefits, while also denying a real human the opportunity to start to contribute to the project.
This discouragement may not be useful because what you call "soulless token prediction machines" have been trained on human (and non-human) data that models human behavior which include concepts such as "grace".
A more pragmatic approach is to use the same concepts in the training data to produce the best results possible. In this instance, deploying and using conceptual techniques such as "grace" would likely increase the chances of a successful outcome. (However one cares to measure success.)
I'll refrain from comments about the bias signaled by the epithet "soulless token prediction machines" except to write that the standoff between organic and inorganic consciousnesses has been explored in art, literature, the computer sciences, etc. and those domains should be consulted when making judgments about inherent differences between humans and non-humans.