The thing is, logically the distribution of opinions or individual situations/beliefs which lead to those opinions, has been baked into the model when the data used to train it was captured, which means that at best and if the entire principle of the thing works (which itself isn’t mathematically proven in any way form or shape) they’re still getting only poll results for the past and which will not actually change beyond some random noise until the next time data is captured and the model is retrained.

It’s like repeatedly using an old picture of a street to make realtime claims about the traffic there.