In the toy example given in the linked page, with the A->B link in place the global optimum (in terms of total minutes spent travelling) is for 1750 cars to tra

In the toy example given in the linked page, with the A->B link in place the global optimum (in terms of total minutes spent travelling) is for 1750 cars to travel Start->A->End for 67.5 minutes of travel, 500 cars to travel Start->A->B->End for 45 minutes of travel and 1750 cars to travel Start->B->End for 67.5 minutes of travel. This gives a total of 258750 minutes spent travelling. If you imagine that one of the Start->A->End travellers switches to Start->A->B->End, we now have 1749 cars travelling Start->A->End for 67.5 minutes, 501 cars travelling Start->A->B->End for 45.01 minutes and 1750 cars travelling Start->B->End for 67.51 minutes. The car that switched improved its own travel time from 67.5 minutes to 45.01 minutes, but the total time spent travelling increased to 258750.01 minutes (one driver saved themselves 22.49 minutes, but cost 2250 other drivers 0.01 minutes each). So yes, if you found yourself in that second state, to improve the global optimum you would have to move to the first state by de-optimising the route of one of those 501 Start->A->B->End drivers.