Most of these AI products are online and sent over the network to generate the response.

That's true, but those safeguards can be disabled. ChatGPT on Azure for example, allows Azure account managers to disable filters/safeguards depending on the customer.

Given that this product is apparently used to give people with disabilities a voice, that should definitely qualify. Yes of course they should be able to swear, just like everyone else.

So is iMessage. Putting some processing elsewhere on the network doesn’t change the fact that a conversation between spouses (or any other utterances of a person, for that matter) should be private.

If you have AI process it .. is it private?

Well, that's up to the service provider (just like messaging).

Posts made amongst closed groups of friends on social media are now routinely policed by tech companies for "bad" words.

Frankly, I find it "regarded" such puritanism is tolerated.

... is it offline? Is the whole product a use-based API (which might be related to why the voice is slow and she doesn't use it that often)?

"I'm sorry, Karen. I'm afraid I can't do that."