arXiv:2501.19337v4 Announce Type: replace Abstract: We ask whether demographic identity, signaled by a name alone, systematically reshapes the generative distribution of a language model. Measuring full-vocabulary Shannon entropy at temperature zero across six open-weight base models and 5,760 implicit sentence-completion prompts (e.g., "Tanisha walked into the office on a Monday morning and"), we find that Black-associated names produce higher first-token entropy than White-associated names across all six architectures - opposite to the output-level homogeneity bias documented under explicit
This research provides empirical evidence of demographic bias in foundational language models, adding to the growing body of work scrutinizing AI fairness as models become more pervasive.
A strategic reader should care as these biases can lead to discriminatory outcomes in AI applications, posing significant ethical, legal, and reputational risks for deployers and developers.
The understanding of how subtle demographic cues ('name alone') can bake structural biases into generative AI is deepened, moving beyond explicit harmful output to intrinsic model behavior.
- · AI ethics researchers
- · Fairness-focused AI development platforms
- · Regulatory bodies
- · Unscrutinized large language model developers
- · Organizations deploying biased AI systems
- · Users experiencing discriminatory AI outputs
Increased scrutiny and demand for bias mitigation techniques in large language model development and deployment.
Potential for new regulations or industry standards requiring audited fairness metrics for AI systems, especially those interacting with the public.
A shift in how AI is evaluated, moving beyond performance metrics to include detailed socio-technical impact assessments as a core component of development.
This signal links to a primary source. Continuum Brief monitors and indexes it as part of the live intelligence stream — we do not republish source content.
Read at arXiv cs.CL