Right, that's the section I was confused by because it was in the context of an experiment trying to use the AI stuff without an Internet connection, which obviously won't work. The article is using the "shared onwards" terminology to refer to at least inference. But the inference part is uninteresting to me, and the data labeling is. The article doesn't really separate those out.
I would figure if there is AI labeling that some things will confuse the system and will be sent to a human. And some things will randomly be sent to a human for error checking. Same thing with Alexa, I figure there's always a low probability chance that anything I say to her will end up reaching a human. She's not always listening as some people fear (the data use would have been detected long ago if she were), but humans occasionally trigger her accidentally--and such errant triggers will be more likely to be sent to a human because they are not going to make sense.