I agree that the summary of this summary (and other stories about this today) make it sound like that, but the more I think of it, that premise seems suspect. Basically, it would be almost constantly streaming audio back to HQ. Now multiply that by 10s (100s?) of thousands of sets across the world, and getting any usable data to improve voice recognition, or parsing which words are intended to be voice control vs. just random talk would take far more computing power and bandwidth than it's worth.
It seems more plausible to interpret the statement to say that while you are issuing voice commands, either via a keyword that can be recognized locally or by pressing a button, THEN any statements that may be unrelated to the functions of the set may be inadvertently picked up and sent to Samsung's (or partners') servers.
Still, even with a favorable interpretation, it seems like an unnecessary, risky, costly "feature" that has only a marginal benefit to customers. Are we so lazy that even pressing a series of buttons takes too much effort? As an accessibility feature, fine, it makes some sense, but it should be turned off by default.