A developer's experiment with Anthropic's Claude models revealed a startling bias in AI randomness. When asked to generate random names 37,500 times, the system repeatedly chose Marcus over 4,300 times, accounting for nearly a quarter of all responses. The most extreme case showed Opus 4.5 returning Marcus 100 out of 100 times with a simple prompt.

This systematic failure of randomness highlights a fundamental challenge in AI development. The experiment tested five different models across dozens of prompt variations, yet nine parameter combinations produced perfectly deterministic output. Even when researchers used elaborate prompts to increase diversity, they discovered new biases rather than true randomness. The team found that random word seeds worked better than random noise for increasing name diversity.

The experiment cost $27.58 in API fees and generated 37,500 JSON responses archived for analysis. These findings matter because developers increasingly rely on AI for tasks requiring genuine randomness, from game design to security applications. The persistent Marcus bias suggests current language models struggle with basic probabilistic tasks, raising questions about their reliability in scenarios where true randomness is essential.