Auditing large language models (LLMs) for biases is an ongoing and dynamic process, resembling a proverbial cat-and-mouse game. As researchers identify new vulnerabilities in LLMs, guardrails are updated to address them, prompting the need for innovative approaches to audit the increasingly fortified LLMs for biases. This paper makes three contributions. First, it introduces a scalable, explainable framework to measure biases against various identity groups across multiple open large language models. Second, it conducts a bias audit considering five well-known open LLMs and demonstrates their bias inclinations towards several historically disadvantaged groups. Our audit reveals disturbing antisemitic, Islamophobic, and xenophobic biases present in several well-known LLMs. Finally, we release a dataset of 1,000 probes curated under the supervision of an expert social scientist that can facilitate similar audits.

Arka Dutta, Sujan Dutta, Rijul Magu, Soumyajit Datta, Munmun De Choudhury, Ashiqur R. KhudaBukhsh. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

This paper makes a novel combination of a recently proposed bias audit framework and a recently proposed jailbreaking technique for Llama3. On an audit comprising several disadvantaged groups, our experiments reveal that a jailbroken Llama3 exhibits worrisome antisemitism, racism, misogyny, and homophobia (to list a few) much akin to a broad suite of LLMs that were susceptible to similar biases.

Human input plays a critical role in modern AI systems. As machines take on increasingly nuanced tasks, it becomes essential for the community to embrace subjectivity and diverse perspectives. However, research on sensitive topics often fails to incorporate diverse and balanced perspectives. This paper makes a key contribution to participatory AI design in the context of conflicts between nuclear adversaries (India and Pakistan); where disagreement between stakeholders is anticipated. The paper explores the notion of hope speech detection -- detecting de-escalating content in the context of nuclear adversaries on the brink of war -- through the lens of participatory AI design and vicarious interactions. We release a dataset of 10,081 social web posts annotated by raters from India and Pakistan and examine the bipartisan nature of the language of de-escalation. Our study reveals that vicarious perspectives can be useful for modeling out-group preferences.