This paper gives a procedure for redistricting which is provably resistant to gerrymandering.

https://arxiv.org/abs/1710.08781

The idea is simple: start with a blank map. The first team draws a districting map. The second team selects one of the districts to “freeze”, erases all the others, and draws their own districts around the frozen district. The players continue passing the map back and forth until all districts have been drawn. Gerrymandering always involves creating some districts which are more favorable to you at the expense of creating some districts which are more favorable to your opponent. If you attempt to gerrymander on your turn, your opponent will punish this by freezing the district which is most favorable to themselves.

We could try and train two AIs to play this game against each other by generating random districts using

https://alarm-redist.org/redist/

If we can train them to get close to optimal play, the resulting districts should be quite fairly drawn. We could confirm that the resulting maps are "fair" by checking to see whether the resulting maps give election results which are close to the statistical average of a sample of randomly drawn maps.