ABSTRACT
The poems the author generates share a few common qualities: they are all presented in game as four-line "extracts" from a larger piece. They all deal with a single overarching theme—a horn, a horse, a politician. When game developer generate poems, their lack of quality is immediately obvious because they are deeply familiar with how language works. It is an effect of uncanny lexis, similar to the uncanny valley—they can make beautiful trees in computer-generated images, but they have not mastered the human face. Familiarity breeds difficulty. Now some insanely clever people at Google—people who are far better at this than the author—have also been teaching computers to write poetry. But instead of procedural generation, they have been using staggeringly complex neural networks. Artificial intelligence (AI) that can learn and adapt, AI that actually understands grammar rather than being told by a programmer to generate text in a way that obeys grammatical rules.
