I usually play on a Standard-sized pangea map with low sealevel and 8 total players (one more than default, to compensate for low sealevel), but I have seen this occur on other settings as well. I reported this problem in the AI feeback thread (in my post "
More Feedback Sample Games"), and attached a saved game with an example of this (see the Dain save attached to that post).
The post-H AI apparently looks around and moves its starting settler to what it calculates as the best nearby starting position (for some definition of nearby). I have seen the AI move its settler up to 3 turns worth of movement to reach a spot; I have no idea what is the actual limit, if there is one, of its relocation.
In my experience this tends to produce, at best, clumping of starting positions as many different AI all converge on a single river. At worst, a group of AIs converge on an area too small to accomodate them all and some of them get stuck standing around because their desired city location is blocked by another city. On a standard sized (non-Erebus) map I have seen up to 3 out of 7 AIs blocked in this way.
Rather than realize that the desired spot is blocked, and recalculate a new desired city spot, the AI will stay put until pushed into a legal city location, and then will immediately found its capital there (no matter how bad that location may actually be). Depending on where their desired location is, the AIs settler may be pushed to a legal city location on the first culture expansion of the blocking city, or not until the second expansion. Because culture expansion can take several turns, and two expansions can take many turns, a civ that gets stuck in this way is severely behind the other civs in the game (and often stuck with a terrible capital location as well). To make matters worse, the AI in this situation doesn't immediately start exploring with its Scout, which means that it rarely gets to explore tribal villages (and also means that other civs tend to get to explore more) - which can further widen the gap, and in some situations can even give one civ a very large advantage over the other civs in the game.
The damage caused by this behavior isn't limited just to the civ(s) that get blocked. Having several weak civs in the game can shift the balance of power very quickly in favor of the strongest civ, as it will generally declare war on and conquer these weak civs with relative ease. I don't mean to say that this AI behavior is a problem (strong civs aggressing against weak civs), but that because of the problems with how the AI handles founding its starting city there exists an unusual disparity between the power level of the strongest and weakest civ in most games. This allows the strongest AI to reach the point at which it can defeat the weakest AI more quickly, which reduces the time until it can begin snowballing in power. Furthermore, the clumping factor tends to produce large empty areas of the map, in which barbarians are able to concentrate. This, and the tendancy of the new barbarians to target a single civ can result in much larger numbers of barbarians attacking a civ as compared to what was seen pre-H.
The large empty areas of the map tend to be expanded into most extensively by the most advanced civs, allowing them to claim the largest area for themselves. They also tend to be in the best position to conquer the barbarian cities that appear in these areas. Weaker civs tend to be more likely to be cut off by the expansion of the stronger civs, and it is not unusual to see one or two AI civs that have more cities than all the rest of the AIs combined - before wars have even broken out.
I realize that I'm speaking in generalizations, but that is necessary when dealing with tendencies and randomness. I'm sure that it is still possible that every AI could start in a location to its liking, and that a dispersed set of starting capitals could all be founded on the first turn, resulting in a fairly balanced game. In general, however, this will not happen, which is unfortunate because I find that the current AI behavior tends to produce games that are less fun to play.
