Pelicans on Bicycles: A Large-Scale Investigation of AI Laboratories

Dylan Castillo conducted an in-depth investigation into the common hypothesis that AI labs intentionally train models to draw pelicans on bicycles. He composed 48 prompts, combining 8 types of animals and 6 types of vehicles, and ran each three times through 7 models: GPT-5.6 Terra, Claude Sonnet 5, Gemini 3.5 Flash, Grok 4.5, Qwen3.7-Max, GLM-5.2, and DeepSeek V4 Pro. For evaluation, GPT-5.6 Luna and Gemini 3.1 Flash-Lite were used. The study found no evidence that images of pelicans on bicycles looked better than other combinations. No lab exceeded expectations in rendering pelicans or bicycles individually, and the 'pelican-bicycle' combination showed no signs of memorization. The GLM-5.2 model showed a slight improvement for this specific pair, but the effect was deemed statistically insignificant.