Token-Level Entropy Reveals Demographic Disparities in Language Models

A new study of six open-weight language models finds that a person’s name alone can systematically alter the distribution of words a model generates, with Black-associated names producing higher first-token entropy than White-associated names across all architectures tested [1][2]. The paper, posted to the arXiv preprint repository and last revised in June 2026, measured full-vocabulary Shannon entropy at temperature zero using 5,760 implicit sentence-completion prompts such as “Tanisha walked into the office on a Monday morning and” [1][2]. Large language models are neural networks trained on vast text corpora to predict the next word in a sequence, and biased or inaccurate training data can make their outputs less reliable [3]. The authors report that Black-associated names always produced greater entropy above identity-neutral baselines than White-associated names, with the difference positive in all six models [2]. Women-associated names co-occurred with lower first-token entropy than men-associated names, with a DL-pooled coefficient of −0.041 and a p-value of .019 [2]. The same names also yielded more homogeneous outputs, a pattern the authors describe as convergent with the output-level homogeneity bias documented in earlier work from 2024 [2]. The pooled coefficient for homogeneous outputs was +0.024, with a p-value below .001 [2]. Race and gender effects were additive [2]. Instruction tuning — a process in which pre-trained models are fine-tuned to follow instructions and behave as assistants [3] — did not attenuate the race gap. In a matched-format comparison, the DL-pooled coefficient for the race gap remained +0.153 [2]. When the researchers replaced names with explicit group labels and ran the same templates, race effects disappeared in 10 of 12 models where the implicit name-based probing had been significant, indicating that the choice of probing methodology is a primary determinant of which distributional structure is recovered [2]. arXiv, where the study appears, is an open-access repository of electronic preprints that are moderated but not peer-reviewed; it was launched in 1991 and now receives roughly 24,000 submissions per month [8]. The platform also hosts arXivLabs, a framework that allows community collaborators to build tools such as citation explorers and code-finders on top of the repository [6][7].