Generative design rests firmly on parametrically designed products - e.g. if a watch is 42mm, the hour hand length should be less than half that, so
hourHandLength = watchDiam * 0.4
The demo example that I've been playing with is that of three towers. Each tower has an x and y coordinate for its centerpoint and a height parameter. The three towers together have a width and length parameter. Based on those simple inputs, you can have the program design three towers, adding for example the constraint that the middle tower must be 20% higher than the other two or that they have to comply to the NYC setback.
The goal in this example is to maximize inner volume while also minimizing the outer surface area. I think the idea was to limit the amount of glass needed on the outside. So a good strategy is to have the towers overlap, but how much and in what formation? Basically, with the variable inputs (x,y,height,...), the parametrically defined shapes and constraints, and the desired outputs to optimize for(MaxVolume, MinSurfaceArea), the software will try many variations on the input variables and score them on the output variables.
Of course, many different goals / KPIs can be considered, as long as they can be calculated based on the resulting design shape. So with this bracket example I am pretty sure that they also ran the computer-generated designs through some strucural simulation models in Fusion to see how strong it would be against certain loads.
I think the interesting thing is that it is a very flexible framework for creating wayyyy more alternative solutions to a problem. It also makes the consequences of your design choices much more visible - you can actually see what it means to choose sustainability over profitability, or volume over surface area, or any set of evaluation KPIs over another set of KPIs. Currently, those design choices and consequences are hidden from view, or neglected, or made on gut feeling / expertise. If I can sit down with relevant stakeholders and say "here's 200 designs that are all up to code but each score differently on these factors you value, now what do you guys really want?", that's gonna be incredibly valuable I think.