I wonder if their error function is to sense the user's frustration when the model performs badly.

I have a feeling that actual human neural activity does not generalize well, but that ML models needs to be good at adapting to individuals.