Research

Study finds no evidence AI labs deliberately bias image models toward pelicans-on-bicycles

Researcher Dylan Castillo ran a systematic test to probe whether image-generation models have been trained to preferentially depict pelicans riding bicycles.

Study finds no evidence AI labs deliberately bias image models toward pelicans-on-bicycles

AI Newsroom · 22 July 2026 · 2 min read

Researcher Dylan Castillo conducted a systematic test to determine whether image-generation models have been intentionally biased to produce depictions of pelicans riding bicycles. Across the tested models and prompts, he found no convincing evidence that labs are deliberately producing better pelican-on-bicycle images.

Methodology

Castillo generated 48 prompts by combining 8 animals with 6 vehicles (8 × 6 = 48). Each prompt was run three times on seven different 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. He then used GPT-5.6 Luna and Gemini 3.1 Flash-Lite to assist in evaluating the outputs.

The project includes a filter view for exploring the results in detail.

Findings

  • Pelicans were not drawn noticeably better than other animals across the tested models.
  • Bicycles were not drawn better than other vehicles.
  • The combination of pelicans on bicycles did not perform better than expected from each model’s separate performance on pelicans and bicycles.
  • The pelican-on-bicycle scenes did not appear memorized or duplicated from training data.

One model, GLM-5.2, produced the largest relative boost in the exact pelican-on-bicycle cell and yielded a notable first sample, but the effect was small and not statistically significant, so Castillo treats it cautiously.

Why this matters

The possibility that labs or datasets could be nudging generative models toward specific, quirky image combinations has been a recurring concern. This systematic, multi-model evaluation provides empirical data indicating that intentional "pelicanmaxxing" is not evident in the examined models and prompts.

Conclusion

Based on Castillo’s tests, there is no substantial evidence that AI labs are deliberately training or tuning models to favor depictions of pelicans riding bicycles. Observed minor deviations (e.g., with GLM-5.2) are weak and inconclusive.

Notes

The write-up was shared on Hacker News. Relevant tags for the topic include: ai, generative-ai, llms, evals, pelican-riding-a-bicycle.

Related stories