Are AI labs pelicanmaxxing?
2 hours ago
- Dylan Castillo conducted a systematic study to investigate whether AI labs are deliberately training models to produce images of pelicans on bicycles (pelicanmaxxing).
- The methodology used 48 prompts (8 animals x 6 vehicles) run three times each through 7 different AI models, with evaluation by GPT-5.6 Luna and Gemini 3.1 Flash-Lite.
- No evidence of pelimaxxing was found: pelicans on bicycles were not drawn any better than other animal-vehicle combinations, and no lab showed a significant advantage.
- GLM-5.2 came closest to showing a boost for the pelican-bicycle prompt, but the effect was small and not statistically significant.
- The findings suggest that the widely discussed 'pelican on a bicycle' benchmark does not reflect deliberate training bias by AI labs.