I would be shocked if an LLM is involved in any step of the process here. I guess the odds aren't zero, but swipe input is well over a decade old at this point. I see it as essentially a solved problem that's entirely doable with standard algorithms - not models, let alone the large language kind. Maybe some model could be used to analyze the context and shift the weight slightly to more probable phrases, but at the core, when you look at inputting single words, it's just checking the letters you swiped over, input timing, turning points and adjacent letters against its dictionary.

according to the article, the blacklist appeared at the same time as the switch over to an LLM based auto complete. By my reading, but I’m not the world’s best reader.

According to TFA the keyboard now uses a transformer based system to handle input and autocorrect.