Privacy-preserving active learning offers a path for heritage language revitalization

A machine learning engineer working with a small Ojibwe language circle found that governance, not compute, was the main obstacle to digitizing teaching materials. Community members raised concerns about where their voices would be stored and whether their annotations could be reused without consent. Heritage languages face scarce data, expert-only labeling, distrustful stakeholders, granular consent requirements, and multiple working languages. To address this, the author proposes a layered design combining federated learning, differential privacy, secure aggregation, and consent-aware data governance. In this setup, active learning queries are scored locally on community devices, so only privatized model updates leave the community.