AI and Pelicanmaxxing: Myth or Reality According to a Rigorous Analysis?

Dylan Castillo conducted an in-depth study to determine whether artificial intelligence laboratories have intentionally trained their models to excel at drawing pelicans on bicycles. This question, while not very scientific, has sparked a debate among experts in the field.

To test this hypothesis, Castillo employed a rigorous methodology involving eight animals and six vehicles, generating a total of 48 distinct prompts. Each prompt was executed three times across seven different AI models, including GPT-5.6 Terra, Claude Sonnet 5, Gemini 3.5 Flash, Grok 4.5, Qwen3.7-Max, and DeepSeek V4 Pro. The results were then evaluated using GPT-5.6 Luna and Gemini 3.1 Flash-Lite.

Analysis Results

Castillo's analysis found no evidence of what he calls "pelicanmaxxing." Pelicans on bicycles do not appear to be drawn any better by AI models than other combinations of animals and vehicles. Furthermore, no laboratory stood out in the representation of these specific scenes.

The results indicate that pelicans are not drawn any better than other animals, and bicycles are not represented any better than other vehicles. Even the combination of the two did not yield significantly better results than what the models already predict individually for pelicans and bicycles.

A Notable Exception

Among the models tested, GLM-5.2 showed a slight performance advantage in the specific task of drawing pelicans on bicycles. Its first sample particularly caught Castillo's attention, although the observed effect was weak and statistically insignificant.

In conclusion, this study challenges the notion that AI models are biased toward specific tasks like pelicanmaxxing, which could influence how these technologies are developed in the future.