Precincts in most places are designed to have roughly the same number of voters. Hence, Benford's law does not work for them. Here's a good explanation, using Chicago precinct data [1].
A good rule of thumb is that if Benford's law does not apply to the legal voter population of your precincts/counties/wards/districts/whatever, or when applied to the total number of registered voters in those areas, then it won't apply to the results of a fair election either.
(This is why if you ask every adult you know their height in inches and then look at the distribution of first digits, it will look nothing at all like a Benford's law distribution. The heights are going to be the sum of two normal distributions (one for men, one for women)).
A good example of that is the county results in Georgia. There is a video going around from a forensic CPA showing the votes by county did not followed Benford's law, but he failed to look at the underlying population.
I've written a little web-based simulation using the Georgia data to make it easy to play with the data to get a feel for why Benford's law does not work there [2].
Another way to tell when a deviation from Benford's law is not an indicator of election fraud is to look at what it would take to make the results fit Benford's law. If you can't find any reasonable way to adjust the votes to make it fit, then the lack of fit was not a fraud indicator. (And if you can find such an adjustment, it also needs to stay a Benford fit if you make minor changes to the voter turnout or the split).
In the particular case of the 2020 US Presidential election, there is another thing about all of the purported proofs of Biden fraud by Benford: cherry picking. There are plenty of places where the precinct vote for Trump does not fit Benford and the Biden vote does. And there are plenty of places where neither the Biden nor Trump vote fits Benford's law.
As you will see if you play with my Georgia simulator, when you have a situation where Benford's law does not apply even a small change in voter turnout or in a candidate's share can cause large changes in the first digit distribution. Because of this, and the large number of cities and counties we have in the US where the underlying population does not follow Benford, there is essentially a large randomness in the first digit distributions.
[1] https://www.youtube.com/watch?v=etx0k1nLn78
[2] https://github.com/tzs/georgia-benford-simulation