The best failure mode is none of those. Using your example, the best one is the one that finds the closest approximate word that isn't a slur. That would solve

The best failure mode is none of those. Using your example, the best one is the one that finds the closest approximate word that isn't a slur. That would solve both problems at once, because there is a really strong chance that the first closest guess that isn't a slur is the actual correct word that was said there. And in case it picks an incorrect word, it would have picked it either way and it would have been just as wrong. So if the voice said "Because lucky you, that's why", in case youtube autotranscript thinks it was "fuck" and not "lucky", it would either default to "lucky" (which is good) or another closest non-slur word like "mucky" or "rocky". And, in my opinion, this is way better than putting [censored] or any other workaround out of those you mentioned, because all of them have zero chance to end up as the correct word (in case the slur was detected incorrectly), while with the approach I described, it is very possible.