Moreover, a site like Facebook is attractive NOT because of its design (mainly) but because of the service it provides. So a Facebook designed clone will have no interest, even if it looks like Facebook.
To me, it looks like it's just the wrong tool for the problem to be solved
Neural nets can approximate any function as accurately as you'd like. Blindly shoving whole sites prelabled by their creativity into a convnet seems likely to end poorly, but that doesn't mean a better training method can't be applied which more faithfully represents the problem domain (no clue what the inside of this particular tool looks like, just commenting on the general infeasibility claim).
> So a Facebook designed clone will have no interest, even if it looks like Facebook.
Doesn't this site just claim that appearances also matter, not that a polished turd will be successful?
I agree: the main problem of "AI" is the knowledge base used. And the biais associated. Which mean that you can correct any biais by changing the reference... but introducing other biaises.
It would be possible to use such a tool either to focus on smaller tasks (let's say: focusing only on photo gallery of professional photographer websites) and be more specific (but less creative) or to grow the knowledge base to include paintings, architecture, object design to exact more general design principles (gestalt and so...)
In the end, the "AI" will be to create new "mixes" of different already used concepts... but I don't see how it would be able to create new concepts. The "AI" will - as much as I understand the technology - stay inside the space defined by it's knowledge base. If all the website of the knowledge base have only white or black background, the "AI" can't "think" to use a green background, because it doesn't have any inference mecanism to think of the background color as any color. It is limited to the background that was fed.
"AI" that does what we want is limited to whatever rules we impose on it. For a lot of problems the most efficient way to impose rules is to provide a set of samples and interpolate, but if we have some way to meaningfully define creativity (which I don't think will be feasible in general any time in the near future, maybe ever) then we can produce an architecture which matches that definition (and if we're hung up on the generative portion of that, a trivial though expensive way to accomplish generation is to enumerate outputs and check if they match our definition for creativity).
"AI" isn't limited to the samples it's given; it's limited to the biases we impose. We can explicitly impose a bias that says hue matters if we so desire.
However, the main problem with creativity is not reducing the knowledge base, it's finding new knowledges to extend the knowledge space consistently. Sometimes enumeration or automatic generation can help... but in that case, it's "only" a fixed set of meta-rules
It looks like everybody is more or more tired of all the websites looking the same because of the "bootstrap"/"material" design. Your tool will only score any website against the tiring majority, so in fact I'm really not sure that it's such a good idea.
I think that a better - but really far more difficult - task would be to different and usable. But you can't achieve this only with "AI" on website: you need to add also some kind of graphic art knowledge and UX principles. Example: instead of considering a color scheme against what is used all over the place (hint: default bootstrap or default material), you should check color scheme against paintings or photographies, and score better what is "beautiful" AND original. Maybe you could add for example some link to kulers and that kind or color scheme design tool. THAT would make it a VALUABLE tool, because it would help the website to stand out and be usable...
Just my 2 cents