Given that that would be opening the manufacturers up to an _extraordinary_ liability if it was ever discovered; that many individuals are highly motivated to discover this behaviour; that it would be possible to do so fairly easily with network monitoring that's well-within the capabilities of many HN readers; and that to the best of my knowledge no-one has ever made such a claim - that's an extraordinary claim that requires extraordinary evidence. If it is pretty well-established, you won't have any trouble providing an example?

You may be confusing this with the fact that recordings are transcribed to improve accuracy[0]. That's well-known and acknowledged! And for folks who _never_ want anyone else to read what they say to voice assistants, that is indeed a problem. But, again - the overwhelming weight of evidence suggests that the only data that is actually transmitted by these devices is when it believes[1] it has cause to do so. I have never seen one iota of evidence to suggest that smart speakers are designed to listen-and-transmit at times other than when there is a reasonable expectation that customers want them to do so.

[0] https://www.rd.com/article/is-alexa-really-always-listening/, "What are the privacy concerns", for instance

[1] yes, false-wakes happen. That's also an issue. But it doesn't support your claim of intentional ("randomly" - I'm assuming you mean this colloquially for "with no justification", rather than literally "in an unpredictable fashion") transmission.