I am almost sure they use AI to generate subtitles too. Some subtitles have sentences that really look like automatic transcription errors.
Probably an art form that's going the way of the dodo, but nevertheless...
Timing particularly! You'll often get a mystery where the detective says "And the killer is … [long, dramatic pause to survey the room] … the butler!", but the subtitles, displayed at the beginning of the sentence, just say "And the killer is the butler!"
Then there's other shows I watch on Netflix will do that exact thing you mentioned - spoiling the punchline because the subtitle appears before a single word is spoken.
Translation is interpretation to start with, but subtitling definitely adds constraints around what works well that don't exist for, say, generic prose in a purely written context. Similar to (but different specific constraints) to translation for dubs.
Why would I want english subs for english audio to be anything than the lines coming from actors' mouths. English subs are the only language you're gonna get from english audio transcription regardless of the underlying technology.