The elephant in the room is the likely voters who either (1) refuse to answer or (2) outright lie about their preferences.
There was the 1982 CA governor's race, in which Bradley (D) was leading the polls, but Deukmejian (R) won.
There was the 1989 VA governor's race, in which Wilder (D) was leading by 10 in the polls, but ended up winning by a mere 0.4% over Coleman (R).
There was the 1992 Presidential election, in which Clinton polled at 49% but ended up winning with just 43%. Perot polled at 14% but ended up with 19%.
And then there was 2016.
People don't always tell the truth to pollsters, and such untruth is not necessarily evenly ideologically split, as the above example show. In particular, people don't want to tell a stranger that they plan to vote for someone who is widely regarded as fascist.
Biden beating Trump by 12 in a poll probably means he will win MI by 3. And, yes, Biden can do that while losing Bay, Macomb, and maybe even Saginaw Counties; his gains in Ingham, Kent, Oakland, Washtenaw, and even Livingston Counties will more than make up for those losses.
Shy Trump voters aren't a thing. They tend to be very boisterous about their support.
In fact, "shy x voters" in general aren't a thing. The primary culprit for election results falling outside the margin of error of polling is due to methodology and sample weighting issues, not dishonest respondents.