Pre-Forgettable Models: Prompt Learning as a Native Mechanism for Unlearning

Unlearning can be built into AI systems from the start rather than added later: The paper proposes designing foundation models with "prompt tokens" that store information about specific classes, allowing instant removal of that knowledge by simply deleting the prompt—without retraining the entire model or needing access to original data.

This approach addresses real regulatory requirements: The framework directly tackles GDPR and similar privacy regulations that require organizations to remove individuals' data from AI systems upon request, offering a practical solution to a growing legal compliance challenge.

Traditional unlearning methods are impractical at scale: Current approaches like retraining, activation editing, or distillation are computationally expensive and slow, making them unsuitable for real-time systems or continuously evolving AI deployments—a problem this method solves.

The method maintains strong privacy protections: Experiments show the framework resists membership inference attacks (where attackers try to determine if specific data was used in training) and prevents knowledge extraction even under adversarial conditions, ensuring forgotten data truly stays forgotten.

Performance on remaining tasks is preserved: The framework successfully maintains predictive accuracy on classes the model should remember while effectively erasing knowledge of classes designated for removal, suggesting a practical balance between privacy and utility.