AWS ML BlogTuesday · July 7, 2026FREE

Teaching models to forget: Selective unlearning with Amazon Nova

AWS announced selective unlearning for Amazon Nova models, a technique that enables the removal of specific data points from a trained model without requiring complete retraining. This approach, detailed in an AWS ML Blog post, allows developers to address data privacy, copyright concerns, or compliance requirements by forgetting particular information. The method leverages a gradient-based approach to update model weights, minimizing the impact on overall model performance. While the post does not specify exact performance benchmarks or model versions, it positions selective unlearning as a practical tool for responsible AI deployment. The feature is available through Amazon SageMaker, integrating with existing workflows. This development reflects ongoing industry efforts to balance model utility with data governance, offering a more targeted alternative to retraining from scratch.