And these are not power that will have intrinsic knowledge or care about things we may understand intuitively: the error rate of a prediction model, false positives/negatives in classification, the difficulty of producing a reliable model of human emotion, how easy it is to build bias into a ML model, etc.
Besides the feedback to the organizer part, there could be all kinds of stats about the meeting, compare to the stats of e.g. a sports game. If you have a pool of sales reps, and you want to help them improve, you can compare the stats: on average you ask 2 questions per call, while the average rep asks 10, etc..
Don't get me wrong, there are a lot of nefarious use cases, but there are definitely positive ones as well.
Maybe the negatives could be mitigated by doing the automated analysis on the client and asking the users if they want to send the feedback, similarly to how crash reports for software are often done.
Also, even if this feature gets implemented, someone will defeat it easily. ( Example: Teachers ask you to keep your camera feed on during the class? OBS + Virtual Camera )
This is one's dangerous. I'm actually finding it hard to think of a use case for this where it isn't abusive and oppressive.