BLASFEMIA MAG

Can AI do “taste”? For a while now, whenever I have a few tokens left over, I’ve been trying to teach “my Claude” “taste”. Technically, what I’m after is being able to set any design task (from a landing page to a poster) and get back a graphically interesting result in a particular style that is defined, but not down to the last detail. The challenging part: unlike code, whose correctness, maintainability and vulnerability are measurable, or at least validatable and optimizable, “taste” works in a far more diffuse way. To make it even harder, the target style is supposed to deliberately work with breaks that, judged purely “objectively”, could be rated as mistakes. At the very least it should be as far away as possible from the “conventionally pretty” of the usual. Much like in language, where a particular style or a sense of humor always lives in the break with the “usual”, not in perfect mastery of it (even if today’s culture machine absorbs the “break” into the norm faster than an Insta reel is forgotten). Why all this: AI is capable of producing perfect-looking UI, at least for the web, when: it’s left completely free and is allowed/asked to work out its own style from nothing but a briefing (at first glance the results here are often impressive) it’s asked to copy a style 1:1 or follow a detailed branding manual It “fails”, in very clear quotation marks, when it comes to grasping a “style” on a somewhat more abstract level and deriving something new from it. That’s my impression at least, and it applies to language as well as to graphic output. Some will interject: “Yet.” And anyway: what does it matter if the two cases above work? Well, my impression so far is simply that the field in the middle, the field of the “vague idea”, is exactly the interesting one. It’s the field where we need to check whether AI will eventually take over here too, in what I’ll call the creative field. We also need to check how far along it is, and whether (or where) it helps or maybe even gets in the way. In all of this, I’m less interested in proving one thing or the other. I’m more interested in finding the altitude at which a creative process has to take place (and graphics are basically just a cipher for a whole range of knowledge and creative work). At that altitude, what comes out at the end isn’t arbitrary and doesn’t scream AI smell. Instead, it a) understands what is one’s own and then b) transfers it to situations that weren’t named beforehand. Thought further, this leads straight back to the very concrete: what does an optimal interface for working with AI look like? Attached is a little sampler of the iterations and test-results so far. My conclusion so far: it is at the very least extremely hard, and I don’t dare speak of success, not even remotely. (Translated from German by Claude)