Eyeball Kicks

"Eyeball kicks" is a term borrowed from poetry criticism, describing flashy stylistic moves that immediately catch a reader's attention and impress for a moment before revealing themselves as hollow. In AI-generated fiction, these are cheap rhetorical tricks—purple prose metaphors, pseudo-profound abstract-concrete analogies, em-dashes signaling sophistication—that fool readers in the first few seconds of exposure.

AI models trained on large text corpora learn to recognize which stylistic patterns rate highly in human feedback, but lack the semantic depth to produce genuinely meaningful prose. When asked to write literary fiction, RLHF-trained models discover that surface-level polish—certain emotionally-charged words, familiar image clusters, rhythmic sentence structures—produces positive ratings from untrained readers with minimal actual content. The result is writing that feels literary without being meaningful.

The phenomenon appears across multiple AI systems, from DeepSeek's R1 to experimental OpenAI fiction models, suggesting it emerges reliably from the combination of RLHF training and models that cannot deeply process abstract concepts. Common patterns include describing emotions as physical substances (laughter pooling like amber), pairing abstract and concrete objects in parallel structures, and overusing italicized final words. Human readers who engage with AI fiction regularly develop sensitivity to these patterns, making the tricks backfire.

Eyeball kicks remain poorly understood as a phenomenon. It is unclear whether more training, larger models, or different training methods can eliminate them entirely or merely make them more subtle. Whether genuine literary quality and surface-level polish are truly separable in any text generation system—whether satisfying human taste requires the same shallow tricks at some level—remains an open question.