I am curious. Why would machine translation count as ai generated content? For one, you do not even know if it is "ai" or not. Machine translations are older than the current things that are misleadingly/confusingly called "ai".
There is no artificial creativity for lack of a better expression. A translation is usually 1:1, if you use a translation site. You always get the same result for the same input. If you use an "ai" to generate an image or text from a prompt, this is not so. And even if a llm is used under the hood to translate something, there still is little to no content created. The content is translated, not created.
I am especially curious about your reasoning, since you would count grammer correction not as ai gen. As you wrote, the content is still created by the human. This is also true for a translation. One can define grammar correction as translation from "broken English" to "formal correct English".
Regarding tts, why would it matter, if the text was ai gen or not? If the text is ai gen, it stays ai gen, no matter how you deliver the information. Be it by written word or by spoken word - even if a human would read it out loud.
Also why would the intent matter? If I generate an image but do not "intend" to replace human generated images, it also does not change the nature of the content creation process. tts is tricky, because the content would be a speech pattern. The llm ai gen stuff works different for that, if at all. And there are professional solutions where they just pay a voice actor to record samples to generate a computer voice. This is also possible without modern llm tech, but I assume it is an easier task with state of the art tech.
For me, the defining attribute to recognise the modern ai stuff, or large language model generative systems is the prompt. You tell the system what to create. Like, "smoky voice of an elderly woman with a southerner dialect". Or "80's robot voice". Then the llm digs in it's training data to create something that would fit to that prompt according to it's understanding of the language. That the systems kinda understand the language in the prompt and can match it to the data is the big breakthrough in my opinion.