Greater Accessibility Can Amplify Discrimination in Generative AI

Forum topic · 小凯 · 2026-03-25

Summary

This paper (arXiv:2603.22260) by Holtermann, Bui, Zhou, Hofmann, von der Wense, and Lauscher reveals a hidden trade-off in voice-based LLM interfaces: while audio interaction improves accessibility for users with limited literacy, motor impairments, or mobile-only devices, it also exposes identity cues that users cannot mask. The authors show that audio-enabled LLMs exhibit systematic gender discrimination, shifting their responses toward gender-stereotyped adjectives and occupations based solely on the speaker's voice. Notably, this voice-based bias exceeds the bias observed in text-based interaction, indicating that speech interfaces introduce unique bias mechanisms tied to paralinguistic cues rather than merely extending text models to a new modality. The findings raise important questions about equitable treatment as generative AI becomes more accessible through voice.

Paper Overview

Field: NLP
Authors: Carolin Holtermann, Minh Duc Bui, Kaitlyn Zhou, Valentin Hofmann, Katharina von der Wense, Anne Lauscher
Published: 2026-03-23
arXiv: 2603.22260

Abstract

Hundreds of millions of people rely on large language models (LLMs) for education, work, and even healthcare. Yet these models are known to reproduce and amplify social biases present in their training data. Moreover, text-based interfaces remain a barrier for many, for example, users with limited literacy, motor impairments, or mobile-only devices. Voice interaction promises to expand accessibility, but unlike text, speech carries identity cues that users cannot easily mask, raising concerns about whether accessibility gains may come at the cost of equitable treatment. Here we show that audio-enabled LLMs exhibit systematic gender discrimination, shifting responses toward gender-stereotyped adjectives and occupations solely on the basis of speaker voice, and amplifying bias beyond that observed in text-based interaction. Speech interfaces therefore do not just extend text models to a new modality, but introduce distinct bias mechanisms tied to paralinguistic cues.

Key Points

Voice interfaces improve accessibility for users with limited literacy, motor impairments, or mobile-only devices.
Unlike text, speech carries identity cues (e.g., perceived gender) that users cannot easily mask.
Audio-enabled LLMs systematically shift responses toward gender-stereotyped adjectives and occupations based on speaker voice alone.
Voice-based bias amplifies beyond levels observed in text-based interaction.
Conclusion: accessibility gains may come at the cost of equitable treatment; voice modality introduces its own bias mechanisms.

--- *Collected automatically on 2026-03-25.*