I'm no fan of Facebook's privacy track record, but this seems like a pretty alarmist story. Accurate transcription is hard. You need hundreds or thousands of hours of training data to get to WERs that consumers will accept on a single accent in a single language. Then you need to do the same for major accent variations (Irish & Indian english, for example). Then you need to do the same again for other languages. You also need to account for big variations in microphone hardware, perhaps bandwidth (which might affect bitrate of the captured audio), and so on and so forth.

Facebook could theoretically have paid voice actors to produce the training data up front to eliminate concerns, and then never tweaked their models based on user data, but the results certainly wouldn't have been as good.

That all said, if they buried the opt-in, they deserve flak.