arXiv:2606.05936v1 Announce Type: new Abstract: Modern language models rely on pretraining filters to remove undesirable content from training corpora and inference-time guardrails to suppress undesirable outputs during deployment. In this paper, we examine how these filtering and moderation decisions produce forms of epistemic erasure and reveal tensions both across automated systems and between these systems and human judgment. We audit four pretraining filters and three inference-time guardrails on Common Crawl sentences containing gender and regional-origin mentions, together with a manual
The proliferation of advanced language models necessitates critical examination of their underlying data and ethical implications, especially as these models become more integrated into societal functions.
Understanding how pretraining filters and guardrails create epistemic injustice is crucial for developing fair and equitable AI systems, impacting trust and regulatory frameworks.
This audit reveals a systematic issue with how AI systems handle sensitive demographic data, potentially shifting development priorities towards more transparent and robust ethical AI practices.
- · AI ethics researchers
- · Regulatory bodies
- · Human rights advocates
- · Auditing firms
- · Developers of unexamined large language models
- · Users impacted by biased AI systems
- · Companies with opaque AI development processes
Increased scrutiny and demand for transparency in AI model development and deployment.
Development of new tools and methodologies for auditing AI systems for bias and epistemic injustice.
Potential for new legal precedents and industry standards related to AI accountability and fairness.
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