Collaborative Research: ReDDDoT Phase 2: Enabling Participatory Privacy Protections for AI Training Data

Artificial intelligence (AI) works by learning from patterns in data. Building AI technologies depends on acquiring personal data for training models. Responsible development of AI as part of public interest technology (PIT) requires building AI that benefits the public interest while safeguarding personal data used to power AI systems. Safeguarding data require tradeoffs between the level of protection provided and the usefulness of the models created with the data. These tradeoffs create a tension that PIT organizations must resolve. This project engages a multi-disciplinary team across sectors in a combination of ethnographic and computational research to develop novel approaches that can support PIT organizations in deploying data safeguards to build AI. The project uses disclosure limitation techniques to protect the privacy of sensitive information in AI training data. Deploying these techniques, including newer techniques like differential privacy (DP), require making tradeoffs that affect stakeholders in the AI lifecycle. For example, strong privacy protection reduces statistical accuracy, which may ultimately reduce the model usefulness. The project will develop novel methods and best practices for navigating these aspects for PIT organizations. The project will: (1) use ethnographic approaches and qualitative inquiry to identify socio-technical decision points and challenges at PIT organizations; (2) create and evaluate novel approaches to participatory engagement of stakeholders in the deployment process; and (3) build software and communication tools for evaluation and transparency of AI systems that use differential privacy. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.