Uploaded June 2026 | Updated September 2026, 3 weeks ago
Provenance Without Surveillance: Privacy Engineering for AI Content Transparency
Sai Prashanth Chandramouli, Sankalp Jain, and Gayathri Ravi, Meta Platforms, Inc.
As regulators increasingly require disclosure of AI-generated and AI-altered content (e.g., the EU AI Act and California SB 942), the industry is converging on provenance standards such as C2PA. But provenance is inherently dual-use: the same metadata that improves transparency and accountability can also expose identity, device, workflow, or other linkable signals, creating new privacy and security risks for users, advertisers, and creators. Critically, not all AI labeling carries equal privacy risk: fully synthetic content with no user input poses different challenges than AI-assisted edits to user-uploaded or camera-captured media. This talk reframes "AI labeling" as a privacy engineering problem: how do we design end-to-end provenance pipelines that satisfy transparency obligations while minimizing personally identifiable information, preventing cross-context linkage, and preserving product usability? We present a practical framework for risk assessment and controls, discussing data minimization, selective disclosure, threat modeling, retention and access policies, and UI/UX choices, and walk through a realistic deployment scenario illustrating trade-offs across the spectrum of AI assistance. Attendees will leave with actionable guidance for building compliant, privacy-preserving transparency systems.
View the full PEPR '26 program at usenix.org/conference/pepr26/program
Provenance Without Surveillance: Privacy Engineering for AI Content Transparency
Sai Prashanth Chandramouli, Sankalp Jain, and Gayathri Ravi, Meta Platforms, Inc.
As regulators increasingly require disclosure of AI-generated and AI-altered content (e.g., the EU AI Act and California SB 942), the industry is converging on provenance standards such as C2PA. But provenance is inherently dual-use: the same metadata that improves transparency and accountability can also expose identity, device, workflow, or other linkable signals, creating new privacy and security risks for users, advertisers, and creators. Critically, not all AI labeling carries equal privacy risk: fully synthetic content with no user input poses different challenges than AI-assisted edits to user-uploaded or camera-captured media. This talk reframes "AI labeling" as a privacy engineering problem: how do we design end-to-end provenance pipelines that satisfy transparency obligations while minimizing personally identifiable information, preventing cross-context linkage, and preserving product usability? We present a practical framework for risk assessment and controls, discussing data minimization, selective disclosure, threat modeling, retention and access policies, and UI/UX choices, and walk through a realistic deployment scenario illustrating trade-offs across the spectrum of AI assistance. Attendees will leave with actionable guidance for building compliant, privacy-preserving transparency systems.
View the full PEPR '26 program at usenix.org/conference/pepr26/program
