GoKawiilBlogger Simon Willison examines why LLM-generated pelican SVGs share a common visual style
Simon Willison analyzed outputs from his recurring 'generate an SVG of a pelican riding a bicycle' benchmark test, noting that SVGs produced by different large language models tend to converge on similar visual characteristics. He compiled original SVGs from his test runs using each model's highest reasoning setting that returned a valid SVG.
GoKawiil's interpretation of the reporting above, not reported fact.
The recurring test has become an informal, if quirky, benchmark for comparing how different models handle spatial reasoning and creative generation tasks. The similarity in outputs across models could suggest shared training data patterns or common limitations in how LLMs translate conceptual prompts into structured vector graphics, though the piece does not establish a definitive cause.
- Willison runs a recurring benchmark prompting LLMs to draw a pelican on a bicycle as an SVG
- Outputs from different models show a recognizable, similar visual style
- The comparison uses each model's highest reasoning setting that successfully returns an SVG
Source: arcfu.com — Aman Rusia, 2026-10-14
Published there as: “Why do LLM pelicans look like this?”
Read the original report → The summary and analysis above are GoKawiil's own, written from reporting by the source above. Facts and quotes belong to the original publisher.