I asked Claude for 37,500 random names, and it can't stop saying Marcus (github.com)

In an intriguing experiment, researchers tested the randomness of Anthropic's language model, Claude, by prompting it to generate 37,500 random names. The results revealed that "Marcus" was overwhelmingly selected as the most common male name, appearing 4,367 times, which accounts for 23.6% of the outputs. Notably, the Opus 4.5 model consistently returned "Marcus" in every single instance—100 out of 100 times—with a simple prompt, highlighting a significant issue with randomness and determinism in AI outputs. The findings prompt a broader discussion on the limitations of language models in producing truly random content. The study discovered that nine different parameter configurations resulted in zero entropy, indicating perfectly deterministic results. While more complex prompts increased name diversity by doubling unique outputs, they also introduced new biases. Additionally, using random word seeds proved more effective than standard random noise for enhancing output variability. The full experiment, along with detailed statistical analyses and a breakdown of API costs, emphasizes the challenges language models face in embracing randomness and the implications this has for their application in creative and diverse content generation.