Silver assimilates poll data into a model that attempts to compensate for bias in the polls. Since some polls are totally fabricated and manipulated to begin with, that’s like measuring things with a rubber ruler and then attempting to infer the actual truth from the resulting mess.
Great analogy. Garbage in, garbage out.
Do they really just make up the numbers? Pull them out of thin air?
Polls can be skewed by taking an unrepresentative sample of the population, and a smart analyst can take that into account.
Of course, even a smart analyst might not spot some bias in the data, but it's not likely that most polls are completely fraudulent.
When Silver got Trump's prospects wrong -- which was most of the time -- was when he ignored the numbers and relied on his own judgement of what was possible.