Ayse Kok Arslan

Ayse Kok Arslan is an AI governance researcher in Silicon Valley affiliated with the Silicon Valley Institute for Humane Technology Systems (SVIHTS), an independent, interdisciplinary institute established in 2026. The institute’s question is how advanced systems stay accountable once they leave the lab and enter real institutions, and its research register runs from AI evaluation and documentation to deployment assurance and public-sector AI adoption. That is the other half of the CalCompute problem: a public compute platform decides who can build AI, and the public agencies and researchers using it then have to decide whether what they built can be trusted.

Her recent work treats governance as an engineering problem. “Advancing AI Governance: Challenges and Opportunities for Reliable Evaluations, Verification, and Regulatory Compliance,” posted to SSRN in March 2025, maps the open technical problems in making AI oversight work: evaluations that are reliable and affordable, ways to verify claims about a model’s properties, and governance of dynamic and multi-agent systems. Earlier papers take on the same question from different sides, including “A Design Framework for Auditing AI” (2020), “An Empirical Model for Validity and Verification of AI Behavior” (2021), “A Benchmark Model for Language Models Towards Increased Transparency” (2022) and “Progressing Towards Participatory AI: A Safety Evaluation Framework.” Her newer work on agentic systems, including “Documentation as Infrastructure for Agentic AI” and “Re-conceptualizing LLMs as Tool-Makers,” argues that documentation should work as operational infrastructure for auditability and traceability rather than as after-the-fact reporting.

A second thread concerns government itself. “An Empirical Model for Exploring AI in Government: Putting Socio-Technological Systems Perspectives into Use” (2021) and “The Government Stack 2.0: A Composable Architecture Theory for Hyperscale” look at how public agencies absorb AI and what architecture lets them do it without losing control of the system. With Alaa Youssef she co-edited Navigating the Intersection of AI Policy, Technology, and Governance (IGI Global, 2025), a 324-page volume on the ethical, legal and socio-economic implications of AI deployment across health care, finance, transportation and national security.

At SVIHTS that research feeds into applied programs: an AI Governance and Deployment Lab working on evaluation methods, documentation systems and institutional readiness models; the Humane AI Access Initiative, focused on AI literacy and lowering barriers to responsible adoption across communities and organizations; an AI governance course built around real deployment failures; and the Humane Technology Salon Series, which convenes researchers, engineers, policymakers and institutional leaders. Her publisher biography also places her at the Berkeley AI Initiative, and she has worked as a researcher and individual contributor at Silicon Valley technology companies including Google and Apple, most recently as a contract researcher on user interaction design at Google.

Her path into the field ran through learning technology. She has worked on research projects for the United Nations, NATO and the European Union, including a design model for computer-based training developed as a NATO case study, taught as adjunct faculty at Bogazici University in Istanbul, and founded Camp Rumi Technology Literacy Group, a grassroots nonprofit providing digital learning services to primary and secondary schools. She has written or edited three academic books, among them Cultural, Behavioral, and Social Considerations in Electronic Collaboration (2016), and more than 70 articles in peer-reviewed journals.

Arslan holds an MSc in technology and learning from the University of Oxford (Kellogg College), where she also carried out doctoral research in human-computer interaction, and an MPhil in technology policy from the University of Cambridge (Wolfson College, 2015). Her through-line is that an AI system is only as governable as the evaluations, records and institutional capacity around it, which is what a public compute platform has to supply alongside the hardware.