A study titled "The One-Word Census" analyzes answer-choice conformity among 44 large language models using a minimal instrument of 31 single-turn prompts. The research characterizes how models converge on specific answers when asked to pick from a large space of equally valid options, such as naming a tree.
- When asked to "pick a word -- any word," 44 models chose "serendipity" 41% of the time.
- Convergence is extreme in some categories, with one answer taking over 80% of all answers in 7 of 31 cases.
- Conformity varies more than fourfold across models, with persona- and community-tuned models being the most divergent.
- Newest mainline flagships are the most conformist, while rankings remain robust to roster composition (leave-one-family-out rho = 0.985).
- The field is more concentrated than people in 18 of 20 shared categories when compared against human category-production norms.
The study provides a public instrument and code to score models by answer-choice surprisal, revealing structured variation in conformity across different model lineages.