What is NLP in AI?

Text has to be turned into numbers before a model can use it. That means splitting it into tokens, normalising it, and representing each document either as counts of its terms or as a dense vector that places similar meanings near each other. Almost all of the quality of the final result is decided in that preparation, not in the classifier that sits at the end of it.

The characteristic failure is domain shift. A classifier fitted on one corpus performs well on text that resembles it and degrades on text that doesn't, without ever announcing that it has. My Fake News Detection project reached 94% accuracy on a held-out set, which is a statement about that dataset — the honest way to read any accuracy figure is to ask what it was measured on.