Programming languages are for the most part deterministic tools that programmers generally understand and can debug. For this audience, the interpreter or compiler is not a "black box" that we only interact with by giving it inputs and seeing what it does.
Although some prompt engineers may have detailed understanding of the AI model and its training corpus, to most people using this kind of AI today the model is a black box, and it's up to them to try things.
Sure, that happens in programming too, but rarely at the level of a programming language, and much more likely at the level of a complex system of poorly documented and buggy interacting parts, such as a long-standing piece of computing infrastructure that has been iterated upon for decades, and now contains a mish-mash of components, many of which were added by people no longer around to explain how they work.