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

A research study evaluating answer-choice conformity across 105 language models from over twenty laboratories demonstrates pronounced output convergence across open-ended queries. When prompted to select a single word from numerous valid possibilities, models frequently converged on identical selections, with 28 of 96 categories exhibiting an 80 percent or higher concentration on a single answer.

Why it matters

This research reveals that diverse language models from distinct developers exhibit severe monocultural convergence when navigating open-ended solution spaces. Organizations relying on multi-model architectures for decision validation face hidden correlated failure modes and homogenized outputs rather than genuine cognitive diversity.

Key insights

• When prompted to 'pick a word' from broad valid option spaces, 105 language models selected the word 'serendipity' 46 percent of the time.
• In 28 of 96 tested categories with many valid one-word choices, a single response accounted for at least 80 percent of all pooled model answers.
• The study evaluated model convergence using 96 single-turn prompts queried eight times per model, quantified via exact-match scoring and answer-choice surprisal metrics.
• The source text cuts off mid-sentence, limiting visibility into the paper's final conclusions or structural causal factors.

Source

arXiv — Computers and Society — https://arxiv.org/abs/2607.12796