Thais Castello Branco distinguishes creative quality from simply producing a correct answer. Design can include relatively objective checks such as contrast and alignment, while aesthetic choices allow reasonable disagreement. The talk identifies three recurring weaknesses in AI-generated design: repetition, poor fit for the intended context and insufficient interpretation of the user’s intent.
Thais Castello Branco describes a study of more than two million historical websites alongside a synthetically generated comparison set. The reported analysis found increasing visual similarity before generative AI and stronger repetition in the generated examples. The proposed measurement approach extracts features such as color, typography and layout, then trains small classifiers to detect recurring patterns. The presentation claims these probes outperformed common language-model judging approaches, but does not provide numerical results in the transcript.
Thais Castello Branco argues that model training alone cannot resolve these problems because users reveal context and preferences during interaction. The proposed application-level approach combines clearer intent, structured brand guidance and checks that keep an agent’s output aligned with its purpose. Creativity is framed as deliberately varying selected conventions while preserving appropriate expectations, rather than merely increasing model randomness.
Watch on YouTube
