Crowding Out The Noise: Algorithmic Collective Action Under Differential Privacy

Differential privacy protects individuals but may undermine collective efforts to influence AI: The paper finds that when AI companies use differentially private models (DP-SGD), groups of users attempting to collectively steer AI behavior through coordinated data modification face significantly higher barriers to success, creating a tension between individual privacy protection and grassroots algorithmic accountability.

Everyday users have a potential tool to influence AI systems: The research identifies "Algorithmic Collective Action" as a mechanism where citizens, workers, or users can deliberately modify the data they share with platforms to influence how AI systems learn and behave, representing a form of democratic participation in AI governance from the bottom-up.

Privacy-enhancing technologies have hidden trade-offs: While differential privacy is increasingly adopted by regulators and companies to protect personal data, this paper demonstrates it has unintended consequences—it may inadvertently reduce the ability of organized groups to challenge or correct algorithmic harms through collective action.

Multiple approaches to responsible AI can conflict with each other: The paper highlights that different strategies for trustworthy AI deployment (privacy protection, algorithm design, regulation) don't always work in harmony; improving one dimension of AI trustworthiness through differential privacy may compromise another dimension (collective accountability and influence).