I used to work in Market Research R&D, in a non-technical role as a project manager. I was deployed to a project that tried to use affective computing (i.e emot

I used to work in Market Research R&D, in a non-technical role as a project manager. I was deployed to a project that tried to use affective computing (i.e emotion recognition) to understand consumer responses to advertisements. It was a total disaster. We'd hook respondents up to a webcam and record their facial expression as they would watch a series of videos. The vendor's emotion recognition machine learning software would then basically assign scores saying that at this second, the viewer expressed xyz emotion. The project failed for 2 reasons - one was that the theoretical link between what expressions people were presenting, and their actual emotions to a particular piece of media was not fully proven - which meant the model output was not particularly helpful from the beginning. Secondly, and this is really important - the model was trained on images of western faces (i.e white people) - and because our target audience - southeast asians - emote very differently a substantial chunk of the output data needed to be trashed (it couldn't process darker faces well, it interpreted a grimace as a smile, etc) So there you have it - this was something I should at least have anticipated - I got in a lot of trouble