Privacy On The Ground

As policymakers and legal privacy folks wade through the challenges of machine unlearning for data deletion, AI systems continue to evolve, creating wider gaps between privacy and AI governance law and regulations and the tech used to respond to it. Jevan Hutson is one of those rare legal professionals with a firm grasp on the shifting sands of AI and machine unlearning from a technical perspective. He understands how unlearning works and how it doesn't, and he connects the dots between the theoretical side of law and the practical sides of AI governance methods like unlearning. In this talk, Hutson explains the different forms or levels of machine unlearning, its caveats and limitations. Plus, he digs into his recent work on what he and his colleagues call AI entanglement, the knotted layers of AI-based elements that live inside AI models, things that make governing these systems more and more complex.

Machine unlearning is a broad category encompassing a variety of techniques intended to remove - or lessen - the presence of sensitive information or concepts such as personal data, intellectual property like branded content and even artistic styles from machine learning models. In the privacy context, machine unlearning is often discussed in relation to the European Union's General Data Protection Regulation (GDPR) and its AI Act, as well as the California Consumer Privacy Act. But the chasm between what some in the policy world might believe machine unlearning can do and the actual technical capabilities of these methods is wide.

This second episode in World Privacy Forum's series on machine unlearning features Kairan Zhao, a PhD candidate at the University of Warwick in the UK focused on unlearning, privacy and AI safety. In this talk she and WPF Deputy Director Kate Kaye discuss key elements of machine unlearning and areas where these techniques are applied. Zhao explains why there could be a long way to go before machine unlearning can be used at scale as an effective AI privacy tool.

The talk in this episode was recorded in March 2026 in Tucson, Arizona at the Winter Conference on Applications of Computer Vision (WACV) where World Privacy Forum held its third WACV privacy tutorial.

Kairan Zhao's related machine unlearning research:

• What Makes Unlearning Hard and What to Do About It, Zhao et al., NeurIPS 2024
• Scalability of Memorization-Based Machine Unlearning, Zhao et al., FITML @ NeurIPS 2024
• Are We Making Progress in Unlearning? Findings from the First NeurIPS Unlearning Competition, Triantafillou et al.

Featured in this episode:

• Kairan Zhao, PhD Candidate, Machine Learning, University of Warwick
• Dr. Yezhou "YZ" Yang, tenured Associate Professor in Computer Science and Engineering in the School of Computing and Augmented Intelligence at Arizona State University
• Jevan Hutson, Director of the Technology Law and Public Policy Clinic, and Acting Assistant Professor at the University of Washington School of Law
• Kate Kaye, Deputy Director of World Privacy Forum

The Privacy on the Ground intro theme features music by Pangal. Episode music is by Maciej Sadowski.

Machine Unlearning is a field of AI research gaining attention in the policy world. It's often discussed in the context of privacy, particularly in relation to the European Union's General Data Protection Regulation (GDPR) and the EU AI Act as well as the California Consumer Privacy Act (CCPA). At World Privacy Forum's tutorial at the Winter Conference on Applications of Computer Vision (WACV) in Tucson, Arizona in March 2026, we aimed to bridge gaps in understanding between the technical academic community, including machine unlearning researchers, and the AI governance, policy, and legal communities. In this episode you'll hear a chat with Dr. Yezhou "YZ" Yang, a tenured associate professor in computer science and engineering in the School of Computing and Augmented Intelligence at Arizona State University; Yang spoke at WPF's tutorial about machine unlearning for concept erasure.

Topics discussed:

• Why machine unlearning for concept erasure isn't just for privacy
• Benchmarks for testing the removal of data - or even the presence of a visual artist's style - from AI models
• Why Yang wants to see more inclusion of technical AI practitioners in the policy conversation

Research discussed:

• EraseFlow: Learning Concept Erasure Policies via GFlowNet-Driven Alignment
• R.A.C.E. : Robust Adversarial Concept Erasure for Secure Text-to-Image Diffusion Model
• WOUAF: Weight Modulation for User Attribution and Fingerprinting in Text-to-Image Diffusion Models

Featured in this episode:

• Dr. Yezhou "YZ" Yang, tenured Associate Professor in Computer Science and Engineering in the School of Computing and Augmented Intelligence at Arizona State University
• Kate Kaye, Deputy Director of World Privacy Forum

The Privacy on the Ground intro theme features music by Pangal. Episode music is by Maciej Sadowski.

The shift to generative AI has exposed the gaps between data and AI governance policy and real world uses and impacts of generative AI.

Jason Tamara Widjaja can see the disconnects first hand from his perch leading the AI research and AI governance team as Executive Director of AI at MSD - the global biopharma company some in the US and Canada know better by the name Merck & Co.

In this talk recorded in August 2025 Tamara Widjaja discusses the reasons why generative AI created new needs for assistance from AI Verify on AI governance tooling and services, about the challenges of scaling AI testing and selecting appropriate metrics, the importance of inspecting a full system in which AI models interact rather than just standalone models, and the potential lack of access to models and systems used in the highest risk scenarios. The Privacy on the Ground intro theme features music by Pangal. Episode music is by Surly Gates.

Featured in this episode: -Jason Tamara Widjaja, Executive Director of AI at MSD (a.k.a. Merck and Co.) -Kate Kaye, Deputy Director of World Privacy Forum

April Chin is co-CEO of AI assessment company Resaro, one of a small group of invite-only "premier" members of Singapore's AI governance program, AI Verify. In this second episode in Privacy on the Ground's series about AI Verify, Chin discusses the evolution of AI Verify, the methods and approaches it has developed and piloted, recent initiatives such as its generative AI-focused global AI Assurance Pilot, and the obstacles that external auditors like Resaro might face when it comes to accessing important components of AI systems needed for meaningful assessment. The interview was recorded in November 2025. The Privacy on the Ground intro theme features music by Pangal. Episode music is by Zefyr. Featured in this episode: -April Chin, Co-CEO of Resaro -Kate Kaye, Deputy Director of World Privacy Forum

Some might consider AI Verify's "building the plane as it flies" approach to AI governance to be rushed or risky; some might question the organization's deep collaboration with industry. But no matter what, it's hard to argue that Singapore's AI Verify project -- and its unique policy sandbox style model -- is influential and offers a real-world example of AI governance that is willing to test and iterate much in the same way that the tech industry itself does. In this first episode of our three-part series on Singapore's AI Verify, listeners will learn about the history and evolution of AI Verify and its key projects, and get a glimpse from insiders into one of its most recent efforts aimed at assessing generative AI applications.

The Privacy on the Ground intro theme features music by Pangal. Episode music is by Gianpaolo Rinaldi.

Featured in this episode:

• April Chin, Co-CEO of Resaro
• Jason Tamara Widjaja, Executive Director of AI at MSD
• Kate Kaye, Deputy Director of World Privacy Forum

In part three of World Privacy Forum's explainer series about C2PA, we'll talk trust. C2PA is not intended to directly determine the trustworthiness of content, but it does have a Trust Model and C2PA metadata are meant to be used as signals for measuring content trustworthiness. In this episode, you'll learn about the C2PA Trust Model process including the Conformance Program, Trust List, and why the absence of C2PA metadata can affect how trustworthiness of content is measured. This episode of Privacy on the Ground features music by Liam Back and Speedtest.

The Privacy on the Ground intro theme features music by Pangal.

Featured in this episode:

• Kate Kaye, Deputy Director of World Privacy Forum

In part three of World Privacy Forum's explainer series about C2PA, we'll talk trust. C2PA is not intended to directly determine the trustworthiness of content, but it does have a Trust Model and C2PA metadata are meant to be used as signals for measuring content trustworthiness. In this episode, you'll learn about the C2PA Trust Model process including the Conformance Program, Trust List, and why the absence of C2PA metadata can affect how trustworthiness of content is measured. This episode of Privacy on the Ground features music by Liam Back and Speedtest. The Privacy on the Ground intro theme features music by Pangal. Featured in this episode:

• Kate Kaye, Deputy Director of World Privacy Forum

In part three of World Privacy Forum's explainer series about C2PA, we'll talk trust. C2PA is not intended to directly determine the trustworthiness of content, but it does have a Trust Model and C2PA metadata are meant to be used as signals for measuring content trustworthiness. In this episode, you'll learn about the C2PA Trust Model process including the Conformance Program, Trust List, and why the absence of C2PA metadata can affect how trustworthiness of content is measured. This episode of Privacy on the Ground features music by Liam Back and Speedtest.

The Privacy on the Ground intro theme features music by Pangal.

Featured in this episode:

• Kate Kaye, Deputy Director of World Privacy Forum

If you're looking for an accessible overview of how C2PA - aka Content Credentials - works technically and how it relates to privacy, identity, and trust, this is it!

In our first episode in this explainer series about C2PA, we shared an overview of how C2PA works, and what it does in relation to privacy, identity, and trust. In this second episode, we dig deeper into Identity in C2PA. You'll learn about what sorts of users want identity to be attached to content metadata – and who doesn't, and about the risks and tradeoffs of identity in relation to C2PA. And you'll hear from Pam Dixon, World Privacy Forum's executive director, about existing digital identity systems, regulations and infrastructure around the world and how they could relate to C2PA. This episode of Privacy on the Ground features music by Liam Back and Speedtest. The Privacy on the Ground intro theme features music by Pangal.

Featured in this episode:

• Kate Kaye, Deputy Director of World Privacy Forum
• Pam Dixon, Executive Director of World Privacy Forum