I've done digging of my own to do the same Google search tricks and found that you can just connect to a websocket and get a live stream of the questions, possible answers as well as every chat message in the feed. OCR not required.

I was also considering using the chat messages to get the answer. People are either going to mostly say the right answer to sound smart or mostly send the wrong answers to try to trick people. Keep track of which is most common and you have an additional datapoint.