Measuring and Solving AI Design Slop

Measuring Design Quality

Thais Castello Branco defines 'slop' as the convergence of AI outputs toward a mediocre mean, characterized by repetition, lack of fit for the specific context, and low user intent. To address this, Taste Labs analyzed over two million websites to identify objective design features like color palettes, typography, and layout. They developed 'probes,' which are small, specialized classifiers that detect individual design characteristics. When these probes are stacked, their combined frequency predicts the likelihood of a design being AI-generated slop more accurately than using an LLM as a judge.

Solving for Intent and Brand Adherence

Beyond measurement, Taste Labs focuses on inference-time interventions to improve output quality. To combat repetition, they are developing a 'creativity API' that forces models to produce outputs that are intentionally out-of-distribution while still respecting domain-specific constraints. This avoids the pitfalls of simply increasing model temperature, which often results in random, context-blind outputs.

To solve for 'lack of fit,' they built a brand extraction system that decomposes a brand into structured, actionable components. This allows agents to adhere to a specific brand identity rather than defaulting to generic styles. For users without an existing brand, the system provides a repository of pre-curated, cohesive brand systems that agents can adopt to ensure high-fidelity results.