Well, I suggest an outright total ban on anything that purports to read people's emotions. Either it doesn't work, in which case it shouldn't be used, or it does work, in which case it shouldn't be used.
Or maybe it more directly tries to sense emotion by using your webcam to look at your facial expression.
That would be acceptable collateral damage, if it couldn't be permitted without opening the door for the creation of systems that used the information against the analyzed people's interests.
I'd be willing to at least consider a carefully crafted exception. The problem being that when you write such an exception, it tends to be awfully easy to introduce loopholes that, in practice, allow using uninformed pseudo-consent, or false consent with no real alternative available, to use information against people.
I.e. my point is that such a ban would have to be very extensive and invasive, with obvious censorship of small, simple segments of code and whole avenues of basic knowledge. Given some data, you can get a crude emotion detector from facial images or text messages - not state of art but somewhat accurate - with something like ten lines of code, with no previous skill on "emotion analysis", just applying generic ML approaches. I can't imagine how such a ban could be implemented, as so many people would still be able to easily make such systems whenever they wanted to, so the ban wouldn't be effective.
Perhaps you could regulate the application of automated decision making to decisions about people and requiring some review-and-override mechanisms (GDPR has some limited aspects of that), but it's a very different area than just banning knowledge and skills that already exist and are relatively widespread.
Like, if you ban reading people's emotions you effectively have to also ban any human interaction.
I suppose the benefit of a human actor is that you can theoretically fine or jail them if they're found to be malicious or sufficiently incompetent.
However, on the other hand, human actors can explain why they're doing the right thing. Even when they are in fact doing the wrong thing. An AI that is broken incomprehensibly can still be determined to be broken. The human actor causing issues can produce very convincing arguments to avoid termination. Also they can bring in donuts every Thursday to stay on the bottom of the termination list.
[And to be clear. I don't trust the technology at all. I just also don't trust the human system either. A system isn't better because innocent people are oppressed by humans instead of by a computer.]
It's also reasonable to ban anything that claims to be better at it than a human.
Although, I do like your comment about scale. If you have a system that's 99% successful, then if you apply it to everyone in the US then you're failing 3 million people. That's a problem.
Of course, your system might be mechanical OR it might just be a group of people each one just "doing their job." From a result oriented point of view, you might end up with a mechanical system that oppresses less people than a people system.
I don't feel good about either one, but I also don't feel good about causing wide scale misery because at least it's people screwing over other people instead of a machine screwing over people.
[Of course it's worth noting that I don't trust the technology at all. It's just that I also don't trust the human solution that it claims it can replace.]
I am aware of a healthcare company which currently has a model in production which alerts healthcare providers if a person displays suicidal intent. They have several confirmed instances deaths being prevented because of the interventions taken due to alerts from their model. While I don't think ML should be used to manipulate people's emotional states, I think this is a case where having a model that can read people's emotions is a good thing.