“Draw a pelican riding a bicycle” is a favourite AI test; a study found no sign labs train for it
Simon Willison asks every new AI to draw a pelican riding a bicycle in code. A study of 48 similar prompts found no sign labs train for it.
Simon Willison, a software developer who blogs about AI tools, gives every new model the same prompt: “Generate an SVG of a pelican riding a bicycle”. SVG is a text format for drawings, so the model has to write code that draws the picture without seeing it. By 26 October 2026 his blog had 155 posts with the tag, the latest for GPT-6.1 Terra. There is no score; people compare the drawings.
Because the test is so well known, people began asking whether labs train their models on it (“pelicanmaxxing”). In July 2026 Dylan Castillo tested 8 animals on 6 vehicles (48 prompts) across 7 models, 3 drawings each, 1,008 in all, and had an AI judge score them from 1 to 5. The pelican on a bicycle ranked 42nd of the 48 prompts. After adjusting for how hard each prompt was, no lab’s models did significantly better on the pelican than on the others.
A separate site, pelicanbenchmark.com, collects pelican drawings submitted by anyone. It says it is unofficial and not linked to Willison, and model details there are self-reported.
What it means
The pelican drawings remain a fair, rough way to see how well a new model draws with code, since one study found no sign of labs tuning for this prompt. It is a quick visual check, not a score, and better pelicans may simply mean better drawing in general.
Caveats
- Castillo’s study used a single AI judge and a budget of about $80. Finding no evidence of training for the test is not proof that nobody does.
- It’s a fun check of drawing-with-code, not a measurement: one prompt, judged by eye.
- Labs could improve SVG drawing in general, which would lift pelicans along with everything else.