A big one is that designing with this process means at every step you are highly dependent on complex software modeling. Instead of using well known design rules to develop your design, you have to run a sophisticated algorithm that probably needs a hefty amount of compute power. Then you have to run FEA on literally every part, again needing lots of resources. That kind of software isn’t cheap and neither are the computer clusters.

Also, you have to put a lot of effort into defining your requirements very precisely and uncovering hidden requirements. For example, if you need a lot of strength in a part do you need that in both directions or only one? If you only need strength in one direction the best solution might be a steel cable, but I don’t know if there algorithms would consider that.

There’s also part integration. Things have to be designed for manufacturing and service (part tolerance stack up, order of operations, tool and hand clearances, etc.)

You also have to carefully consider your validation and what assumptions it makes about the parts that might no longer be true.

In the end, generative design is probably only worth the effort for specific parts or even portions of those parts, not the whole product.

There's also a problem with adding on stuff later. For example you have some brackets to hold something, one made classically with mostly square shapes, the other with this type.

Then you later need to add some way to attach another thing to the bracket. With the classic version there might be the option to just add some threaded holes and it's done, with the AI/topographical version you'll probably have to start "from scratch".