To your last point, I think it still snapshots ever 10 seconds or so and hash what it has and send a it. I was trying to find the other post on this which described it and failed.

I did find this article, which makes a SmartTV look like a surveillance machine, even with the HDMI input used.

Fielding: Smart TVs gather an enormous amount of data about their usage and their immediate environment (including other devices connected to them, such as speakers, consoles and media storage), which is sent back to the manufacturer. Some of this data is used to troubleshoot and improve the device’s software or media services, but much of it is also used to profile the TV’s users—their viewing habits can be used to make inferences about their politics, professional and economic status, gender identity, sexual orientation, religion, ethnic identity, social activity, purchasing potential, values and beliefs, all of which helps advertising networks know who to target with what.

...

Lewis:* *While the concern is apparent, there currently isn’t a concrete example of a TV manufacturer snooping on its users. In 2015, Samsung landed in hot water regarding an unfortunately worded statement on monitoring living room conversations, stumbling into a communications crisis. Samsung acted quickly, cleaning up and re-wording the statement. However, the immediate outcry was a solid indication of fears around public monitoring. Any proven example of Orwellian-esque monitoring would prove catastrophic for the manufacturer involved.

Kelso: A few years ago, according to the FBI, app developers Vizio, LG, and Samsung were caught snooping on viewers. The FTC had to step in and stop them. Also, the CIA and MI5 were able to access information on smart TVs and listen in on private conversations using the camera and microphones on these devices.

Fielding: Surveillance functions and equipment are built in to almost all smart devices and marketed as “features” to make the user’s life easier. Audio recordings generated from the TV listening for its “wake word” are sent back to manufacturers so they can train their speech recognition algorithms. That’s a deliberate design choice, but it can mean that people’s private conversations are revealed to the manufacturer's employees as a result.