After using these systems I think it's plausible: the current ML systems produce valid but incorrect words frequently and since the UI is designed for marketing rather than usability nobody wants to expose low confidence levels or a correction UI. They have enough customers that someone was bound to get a false positive which the system mapped to one of a few constrained choices in a brittle interface tree.
As an example, I've noticed that Siri tends to default to offering to open Notes on my iPhone any time it fails to recognize a query correctly. Alexa on the FireTV is better but still has lots of failure modes which look suspiciously like the default branch of a switch statement where it ends up really off on a common query in a quiet room.
As an example, I've noticed that Siri tends to default to offering to open Notes on my iPhone any time it fails to recognize a query correctly. Alexa on the FireTV is better but still has lots of failure modes which look suspiciously like the default branch of a switch statement where it ends up really off on a common query in a quiet room.