A lot of input methods are probabilistic in their nature. Some other examples are voice to text and even regular digital keyboard input (fat-finger correction that tries to pick the letter you wanted to tap instead of the letter that you actually tapped). So even with your distinction, actually spelling it out with the digital keyboard would not make a difference because the software still has to make a call there. Even if you plugged in a physical keyboard, Apple could still say that the signal passes through its software, which can constitute endorsement, so it should run the checks against that too, just in case. Nothing in your argument stops that.

I fully reject that this makes the situation any better or more justifiable. The job of the software that picks what you have likely just said, swiped or tapped should only concern the given word. If the algorithm decides with 95% certainty that the input was "dictator", that's what it should output. If context is used to better select the candidate word, it should only be done for the purpose of slightly re-weighting the possible options with one transparent algorithm, not for checks against a blacklist.

The reason why I think this is that your line of thinking is unbounded, and there is nothing stopping the next logical move - if Apple can decide 'not to endorse' (representation isn't endorsement, but whatever) views by blacklisting its algorithms from forming certain phrases, why can't they uplink entire paragraphs of your text as you're typing it to dynamically analyze them instead of using a dumb filter, and then decide if the entire text is something they want to allow?