Pranav A

In this paper, I demonstrated the ineffectiveness of ReID datasets that are being deployed in surveillance pipelines.

We audited marketing materials of 100 startups in Germany and discussed the perception of exaggerated claims (AI-washing) with experts and public.

An early account of my doctoral dissertation on critiques of current policymaking related to surveillance based AI.

Content filtering systems disproportionately remove mentions of marginalized groups while failing to catch genuinely harmful material, creating a form of epistemic erasure in language model training and deployment.

This mixed-methods paper surveys and interviews NLP researchers about ACL's anonymity policy removal and quantitatively analyzes how preprint visibility and author recognition bias review scores and quality across institutional tiers.

Having proactive name change policies leads to inclusive publishing with better citation quality.

Understanding the impact of AI requires more than technical expertise, but also listening to the communities affected by these technologies.

We introduce Glitter, an English-German benchmark featuring extended passages with professional translations implementing three gender-fair alternatives. Our experiments reveal significant limitations in state-of-the-art language models, which default to masculine generics and rarely produce gender-fair translations even when explicitly instructed.

We expand Yann LeCun's 'cake that is intelligence' analogy to the full life-cycle of AI systems, from sourcing data to evaluation and distribution. We describe each step's social ramifications and provide actionable recommendations for increased participation in AI discourse.

Empirical comparisons of character-based models against word-based models on common Chinese semantic benchmarks.

Community-led participatory design case study of Queer in AI contributed lessons on decentralization, building community aid, empowering marginalized groups, and critiquing poor participatory practices.

The paper contributes a contextual subword segmentation method along with benchmark datasets that outperformed previous Chinese character conversion approaches by 6 points in accuracy, especially for code-mixing and named entities.

The paper contributes information retrieval ranking functions with heuristics like positional tokenization and graphical error modelling to the problem of cognate detection.