AI slop is measurable—and fixing it requires judgment, not just bigger models

Taste Labs founder Thais Castello Branco argues that as AI generation costs drop, judgment becomes the scarce resource. The core problem is that AI outputs collapse to the mean, producing repetitive, context-agnostic slop. Their research on 2M websites shows that slop is measurable through structural features like color palettes, typography, and

layout—detectable by simple classifiers better than LLM-as-judge. They advocate a two-pronged approach: improve models via post-training data and RL, but equally important, fix inference-time behavior. Repetition can be broken by a creativity API that intentionally generates out-of-distribution results without random noise. Lack of fit is addressed by a brand API that

extracts structured guidelines from existing brands, allowing agents to maintain consistency. The biggest tradeoff is that taste is subjective—no one-size-fits-all solution. The brand API works for existing brands but struggles for novel contexts. And while measurement helps, it doesn't guarantee creativity. Teams must decide whether to invest in model-layer perfection

or accept inference-time guardrails as the faster path to quality. For engineering leaders, the signal is clear: the bottleneck is shifting from model capability to governance and verification of AI outputs. Building structured context layers—like brand definition, creativity constraints, and slop detectors—may be a higher-leverage investment than chasing ever-bigger models.