Summary
Thais Castello Branco, founder of Taste Labs (recently out of stealth), presents research-backed approaches to quantifying and eliminating AI design slop. The company works on two tracks: collaborating with frontier AI labs on model evaluation and post-training data for design quality, and building application-layer tools to help companies avoid generic AI-generated aesthetics without touching the model layer.
The empirical centerpiece of the talk is an analysis of over two million websites spanning a decade, comparing human-designed sites to AI-generated ones to identify quantifiable slop markers. Taste Labs extracted structured design features — color palettes, typography choices, layout patterns — and trained lightweight probe classifiers to detect AI-generated homogeneity. Combining multiple probes achieved high prediction accuracy, effectively giving the team a measurable definition of slop.
Castello Branco draws a conceptual distinction between design elements that become near-deterministic when properly specified (contrast, alignment, color selection) versus those that require handling genuine expert disagreement (aesthetics, originality). Different training strategies apply to each. The broader argument is that the same rigor applied to math and coding benchmarks can be brought to subjective creative domains — and that solving slop requires measuring it first. The talk is relevant for AI researchers working on creative model capabilities, application developers building design tools, and anyone thinking about how to preserve stylistic diversity as AI-generated content scales.
📺 Source: AI Engineer · Published September 10, 2026
🏷️ Format: Deep Dive
