I work at a self-driving car company and we observed a similar problem when we did some off-road testing on dirt tracks. The cars were too precise and they were cutting deep ruts into the soil. We too solved it by adding a pseudo-random offset to the track.

Before the current wave of automation there was a previous technology to automate buses using optical sensing and lines in the road which had the same issue.

I believe Google Maps adds a bit of a rng in which route it will recommend when two otherwise similar in distance/time. Obviously the traffic input also affects this, but that's a slower feedback mechanism; better to distribute the cars all leaving the airport for downtown across the 2-3 possible routes upfront rather than dumping them all onto route A until it's a jam and then all onto route B until it's a jam, etc.