If I ask a computer to make a bracket lighter, I have already made several decisions. I chose a bracket rather than a different assembly. I decided which loads matter. I decided that weight deserves optimization and that some combination of cost, appearance and manufacturing difficulty is acceptable. The computer hasn't removed those judgments. It has made them part of the input.
Maurice Conti's talk about intuitive AI and generative design is a good starting point for thinking about this shift. Instead of drawing one object and asking software to evaluate it, a designer describes conditions and explores forms produced within them. The author's work moves partly from shaping an answer to shaping the space of possible answers.
Imagine the bracket has two mounting holes, a region that must remain empty and a load applied at an angle. The mounting surfaces are fixed. Material can be removed elsewhere. A search process evaluates candidates against the modeled load and changes the distribution of material. Thin branches and unexpected voids may emerge because the system is responding to stress rather than to the visual conventions of a familiar bracket.
That doesn't make the resulting form automatically useful. Perhaps the assumed force is wrong. Perhaps a technician needs to reach a bolt with a tool. Perhaps the attractive geometry can be manufactured only through a process that is impractical for the production quantity. These are not irritating details added after the real design. They belong in the problem definition.
I find the distinction between a constraint and a preference especially helpful. A part must fit inside an enclosure. That's a constraint. Lower weight is desirable. That's a preference. If the optimizer treats both as interchangeable numbers in one score, it can trade a small violation of something essential for a large improvement in something optional. A mathematically strong result may then be unusable.
There is also rarely a single best candidate. One version may be lighter but expensive to manufacture. Another may use more material while requiring fewer operations. Showing that tradeoff gives the designer useful information. Hiding it behind one impressive image makes the system harder to question.
The same reasoning applies outside mechanical engineering. A procedural environment generator might preserve walking paths while varying room shapes. A layout tool might protect reading order while exploring different arrangements. In both cases, the hard part is deciding what must remain true and how to recognize a failure.
The promise I take from generative design is a larger, more inspectable set of alternatives. I don't want a machine to produce something strange and call the strangeness intelligence. I want it to expose a possibility I wouldn't have drawn, then give me enough information to decide whether it belongs in the world.