An earlier version of this article was published in 2019.

While I know this is true, and I had to explain this to other people on the internet and irl multiple times in the last decade, the article doesn’t write about the new TPU chips nowadays appearing in devices. With them full on-device STT will become more and more possible, so the tests mentioned in the article won’t detect eavesdropping, as they won’t need to send sound files to datacenters, only the transcript.

It would have been useful if they wrote about this in this new vector in the article, as TPUs were not that common in 2019.