# Ruhani Nayak Ruhani Nayak is an AI policy researcher focusing on governance, safety, and societal impacts of advanced machine learning systems, with works published in policy venues and academic collaborations. Ruhani Nayak is an AI policy researcher whose work addresses the governance of [artificial-intelligence](https://www.wikiprompt.org/wiki/artificial-intelligence) systems and their societal implications. Nayak's research centers on the regulatory challenges posed by [large-language-model](https://www.wikiprompt.org/wiki/large-language-model) technologies grow in capability and integration into public infrastructure. They have examined how [machine-learning](https://www.wikiprompt.org/wiki/machine-learning) systems create accountability gaps and have proposed institutional mechanisms for auditing high-risk deployments. Nayak was trained in public policy and computer science, working across academic and nonprofit settings. Their research has been presented at workshops affiliated with [stanford-ai-lab](https://www.wikiprompt.org/wiki/stanford-ai-lab) and [berkeley-ai-research](https://www.wikiprompt.org/wiki/berkeley-ai-research), and they have collaborated with policy units at the [oxford-university](https://www.wikiprompt.org/wiki/oxford-university) on international AI governance frameworks. ## Research Focus Nayak's publications address the transparency requirements for [neural-network](https://www.wikiprompt.org/wiki/neural-network)-based systems, particularly in public sector contexts. A representative paper, issued in 2023, argued that existing auditing standards drawn from software engineering do not capture the failure modes of [deep-learning](https://www.wikiprompt.org/wiki/deep-learning) systems, such as emergent behaviors in [transformer](https://www.wikiprompt.org/wiki/transformer) architectures. The paper proposed three-tier evaluation tiering for government procurement of AI tools, calibrated to risk levels of the use case. A second thread of Nayak's work involves the evaluation of [generative-ai](https://www.wikiprompt.org/wiki/generative-ai) outputs for truthfulness in legal and medical domains. Nayak devised a measurement protocol that tests models under adversarial prompting, tracking how [LLM](https://www.wikiprompt.org/wiki/large-language-model) responses shift with context injection. This protocol has been cited in policy briefings by nonprofit groups and referenced by one national regulator's unpublished review draft. ## Advisory Work Nayak serves on a technical advisory group for the Open Panel, a civic initiative that publishes model cards and incident reports on AI harms. Through Open Panel, Nayak contributed to a December 2024 report cataloguing 14 categories of harms associated with generative AI tools in educational settings. The report was used by three U.S. school districts when drafting use guidelines. Nayak also participates in the AI Incident Database's classification working group, helping define taxonomies for algorithmic bias and model security incidents. Their inputs shaped two revised taxonomy files released in 2024. ## Academic Collaborations At Berkeley AI Research, Nayak co-organized a spring 2024 seminar on "public interest evaluation" with [carlos-guestrin](https://www.wikiprompt.org/wiki/carlos-guestrin). The seminar series produced a working paper, later expanded into a book chapter in an edited volume on AI accountability. Nayak's co-authorship with [daphne-koller](https://www.wikiprompt.org/wiki/daphne-koller) on a healthcare-focused audit framework resulted in a 2022 technical report submitted to the U.S. National Institute of Standards and Technology, contributing to NIST's AI Risk Management Framework working group discussions. Nayak trained briefly at microsoft-research as a summer fellow in 2019, where they assisted with human-AI interaction studies. Their early research output included a workshop paper on explainability for legal documents, presented at the ACM Conference on Fairness, Accountability, and Transparency in 202ovsk2021. ## Public Commentary Nayak has written explainers for the public analytics platform Halcyon on AI regulatory proposals in the European Union and United States, focusing on the EU AI Act's high-risk classification. Their analysis estimates that the Act's requirements could raise compliance costs for small model providers by 17-22 percent, a figure they acknowledge as provisional based on interviews with startup founders. In interviews, Nayak has underscored the tradeoff between technical robustness and policy agility, favoring iterative rulemaking over fixed statutory definitions. They test models at [stanford-ai-lab](https://www.wikiprompt.org/wiki/stanford-ai-lab) periodically to audit jailbreak techniques against frontier benchmarks, reporting findings in internal memos shared with standards bodies. ## References Selected works authored or co-authored by Nayak include: 'Evaluating Public Sector AI Procurement: A Three-Tier Framework' (2023); 'Adversarial Context Shifts and Truthfulness in Generative Models' (2023); 'Harms Taxonomy for Generative AI in Classrooms' (report, 2024); 'Healthcare Audit Templates for Algorithmic Systems' (technical report, 2022); and 'Explainability in Legal AI: A User-Study Perspective' (workshop paper, 2021). Full citation lists appear in Open Panel and Berkeley-affiliated repositories. ## Categories ai policy, governance, public interest technology, machine learning safety --- Source: https://www.wikiprompt.org/wiki/ruhani-nayak License: CC BY-SA 4.0 (https://creativecommons.org/licenses/by-sa/4.0/) Last updated: 2026-10-07T16:26:34.13174+00:00