Kimi K3.1 tops Contra Labs slop detector
Contra Labs evaluated frontier AI models on real-world landing page tasks using its Slop Detector metric, which measures stereotypical AI design patterns on a 0 to 100 scale. Moonshot AI's Kimi K3.1 Pro recorded the highest slop score at 66.4 due to formulaic italic headings and numbered cards, while DeepSeek V4 Pro scored an anomalous 1.2 because its unfinished outputs lacked typical AI styling.
The Slop Detector exposes a harsh reality of generative web development: frontier models often substitute formulaic visual tropes for genuine design taste. Paradoxically, broken code registers as more 'human' to classifiers simply because human designers fail differently than generative LLMs.
- –Kimi's predictable aesthetic crutches: Kimi K3.1 Pro's 66.4 slop score stems from heavy over-reliance on prompt tropes, with 51% of its pages using italicized headline words and 29% forcing numbered section indices (01, 02)—patterns virtually absent from human designer work.
- –The DeepSeek anomaly: DeepSeek V4 Pro scored 1.2—nearly matching human reference designs (0.9)—not due to taste, but because incomplete and broken HTML rollouts don't exhibit the polished, cookie-cutter component cards common to leading models.
- –Lab-specific design fingerprints: Slop scores cluster by model provider rather than coding capability, with Gemini models heavily favoring boxed card layouts (up to 97%), Claude models clustering in the mid-30s, and OpenAI's GPT models hovering between 41 and 53.
- –Implications for UI agents: For developers building autonomous frontend agents, eliminating the 'AI look' requires explicit negative constraints against generic layout boilerplate like tiny spaced uppercase subheadings and decorative card wrappers.
DISCOVERED
18m ago
2026-11-22
PUBLISHED
48m ago
2026-11-22
RELEVANCE
AUTHOR
AdamHoltererer