モデルに忘れさせる：Amazon Novaによる選択的アンラーニング Teaching models to forget: Selective unlearning with Amazon Nova

Amazon Web Services has published guidance on applying selective machine unlearning to its Amazon Nova foundation models, a technique that aims to remove the influence of specific training data from a model without the cost of retraining it from scratch. The approach matters because organizations increasingly need to comply with data privacy regulations, honor deletion requests, and strip harmful or copyrighted material from systems that were expensive to build in the first place.

Machine unlearning refers to methods that make a trained model behave as though certain data had never been part of its training set. Traditionally, the only guaranteed way to achieve this was full retraining, which for large foundation models can require enormous compute budgets, extended timelines, and significant energy consumption. Selective unlearning instead targets only the parameters or behaviors associated with the data to be forgotten, leaving the rest of the model's capabilities intact. The AWS Machine Learning Blog frames this as a practical middle ground between doing nothing and rebuilding a model whenever a removal obligation arises.

The regulatory context is a central driver. Frameworks such as the European Union's General Data Protection Regulation include a right to erasure, often called the right to be forgotten, and comparable provisions appear in laws like the California Consumer Privacy Act. When personal data has been absorbed into a model's weights, deleting it from the underlying dataset does not necessarily remove its imprint from the model's outputs. Unlearning techniques are presented as a way to address that gap, though it is worth noting that the field is still maturing and that verifying complete removal remains technically difficult.

Amazon Nova is the family of foundation models Amazon introduced through Amazon Bedrock, spanning text-focused tiers such as Nova Micro, Lite, and Pro, alongside multimodal variants for image and video generation. Because these models are offered as customizable building blocks, customers can fine-tune them on proprietary data, which in turn creates the possibility that sensitive or unwanted information becomes embedded during that customization. Selective unlearning is positioned as a governance tool for teams that have already invested in tailoring Nova to their domain and later need to reverse specific inclusions.

On the technical side, unlearning methods generally work by adjusting model parameters to counteract what was learned from a target set of examples. Common strategies described in the research literature include gradient ascent on the data to be forgotten, fine-tuning against curated replacement data, and influence-based methods that estimate how particular training points shaped the model. Parameter-efficient techniques such as low-rank adaptation are often paired with these approaches so that changes remain localized and reversible. A recurring challenge is preserving general performance: aggressive unlearning can degrade accuracy on unrelated tasks, a phenomenon sometimes described as catastrophic forgetting, so implementations typically balance the strength of removal against the retention of desired knowledge.

The blog situates unlearning within a broader set of AWS tools for responsible model deployment. Amazon Bedrock Guardrails, for example, filters inputs and outputs at inference time to block disallowed content, and content moderation services can screen material before it ever reaches a model. Unlearning is different in kind because it modifies the model itself rather than wrapping it in runtime controls, and the two approaches appear complementary rather than interchangeable. Guardrails can suppress an unwanted response, while unlearning targets the underlying representation that produced it.

Industry interest in this area has grown alongside scrutiny of how large models memorize training data. Researchers have documented cases where models can be prompted to reproduce verbatim text or personal details, and academic competitions have been organized specifically around evaluating unlearning methods. Against that backdrop, a cloud provider offering documented workflows for its own model family is a notable step, even if the guarantees such methods provide are probabilistic rather than absolute.

Organizations evaluating these techniques will likely need to weigh several factors, including how to measure whether forgetting was successful, how much task performance is sacrificed, and how the process fits into existing audit and compliance records. The AWS material appears aimed at practitioners who already operate Nova-based systems and want a repeatable, lower-cost alternative to retraining when data must be withdrawn. As with any emerging capability, results are likely to depend heavily on the specific data, model tier, and validation methods a team applies.