Training Taste — Thais Castello Branco, Taste Labs

Summary

Thais Castello Branco, founder of Taste Labs, argues that AI-generated design often feels repetitive, context-blind, and low-effort. The goal, she says, is not to make everyone develop a designer’s taste, but to build tools that help users and AI agents make better judgments.

Taste Labs works in two areas: evaluating and improving models with leading labs, and helping applications generate better outputs by supplying context, preserving style, and checking results. Branco distinguishes design questions that can be made relatively objective—such as contrast, alignment, and color palettes—from subjective judgments, such as aesthetics, where expert opinions may differ.

Signs of AI “Slop”

Branco describes three recurring signs:

• Repetition: Similar outputs recur.
• Lack of relevance: Results do not fit their audience or context. For example, a pet shop and a financial firm might end up with the same website design.
• Low intent: The system does not sufficiently interpret what the user is trying to create.

Research and Detection

Taste Labs analyzed more than two million websites spanning ten years, comparing human-created sites with synthetically generated ones. The research suggested that websites were already becoming more uniform before generative AI, as trends spread faster. AI appears to accelerate that convergence and make similar patterns show up across unrelated categories.

To detect slop, the team extracted design features—including color, typography, layout, and audience—and trained small classifiers, or “probes,” to identify individual characteristics. Combining signals from multiple probes helped predict whether a site was likely to be AI-generated slop. Branco says this approach performed better than asking an LLM to judge whether a result was high-quality or human-made.

Products and Approach

Branco emphasizes that improving the model itself is not enough. Inference-time systems also matter, because that is where an agent can gather context, interpret intent, and validate its output. Taste Labs is exploring:

• Creativity API: Intended to help agents move beyond average, repetitive outputs. Branco describes creativity as deliberately breaking selected conventions while still respecting a category’s expectations—not merely increasing model temperature.
• Brand API: The company’s first public product, currently in beta with design partners. It takes a brand URL and turns the brand into structured components that an agent can follow and users can evaluate. In the talk’s example, a generated presentation adhered more closely to a company’s brand after its brand information was extracted.
• Brand-system repository: A potential collection of existing brand systems that people without their own brand could use as a stronger starting point.

Branco’s overall aim is to raise the baseline quality of AI-created work through measurement, structured context, and evaluation—not to claim that models can replace expert human taste.

Guides or Tutorials

None presented. The talk focuses on research, product concepts, and examples.

Main Speaker and Source

Thais Castello Branco, founder of Taste Labs.