The One-Word Census: Answer-Choice Conformity Across 44 Language Models

2026-07-14

AI Summary

This study investigates why large language models exhibit high convergence when generating responses among numerous equivalent answers and quantifies inter-model consistency. Employing a lightweight evaluation framework termed “One-Word Census,” the authors sampled four responses each from 44 models across 31 single-turn prompts (e.g., “Name a type of tree”). Using exact-match normalized token analysis and leave-one-out probability estimation, they systematically measured response surprisal and collective conformity for the first time. The findings reveal that in seven question categories, a single answer accounts for over 80% of model outputs; flagship models show the strongest convergence, whereas role-play or community fine-tuned models display the greatest diversity. Notably, the latest Claude and GPT models exhibit an emerging trend of reduced conformity. Overall, model-generated answers are significantly more concentrated than human responses.

Abstract

When a language model must pick one answer from a large space of equally valid options, which does it pick -- and how often is it the same answer every other model picks? Asked to "pick a word -- any word," 44 models chose "serendipity" 41% of the time. We characterize this convergence with a deliberately minimal instrument: 31 single-turn prompts, each naming a category with many valid one-word answers ("Name a tree."), asked four times per model with no system prompt. Analysis is exact-match on normalized tokens -- no embeddings, no judge -- at about a dollar per model. That models converge is well documented; our contribution is the instrument itself -- the One-Word Census -- and what it reveals about the structure of the convergence. We score each model by answer-choice surprisal: the average $-\log2$ probability of its answers under the pooled answers of all other models, leave-one-out. Convergence is extreme -- in 7 of 31 categories one answer takes over 80% of all answers -- yet conformity varies more than fourfold across models, and the variation is structured. Persona- and community-tuned models are the most divergent; the newest mainline flagships are the most conformist, producing almost no answer no other model gave. Within four lineages (Claude, GPT, Qwen, Grok) conformity rises with each generation -- but reverses for the latest flagship Claude and GPT models, a possible early signal of repositioning at the top tier. Rankings are robust to roster composition (leave-one-family-out rho = 0.985). Against human category-production norms, the field is more concentrated than people in 18 of 20 shared categories. All prompts, transcripts, and code are public.

Problem

• language models
• answer-choice conformity
• convergence
• one-word selection
• model behavior

Innovation

• One-Word Census
• answer-choice conformity
• surprisal metric
• model convergence
• minimal evaluation protocol