Operationalising the Right to be Forgotten in LLMs: A Lightweight Sequential Unlearning Framework for Privacy-Aligned Deployment in Politically Sensitive Environments

Researchers introduce a sequential unlearning framework that enables Large Language Models to forget sensitive data while maintaining performance, addressing GDPR compliance and the Right to be Forgotten in politically sensitive deployments. The method stabilizes general capabilities through positive fine-tuning before selectively suppressing designated patterns, demonstrating effectiveness on the SemEval-2025 benchmark with minimal accuracy degradation.

The emergence of this unlearning framework addresses a critical gap between regulatory requirements and practical AI deployment. GDPR's Right to be Forgotten obligates organizations to erase personal data upon request, yet LLMs trained on internet-scale datasets present unprecedented technical challenges for compliance. This research bridges that gap by proposing a two-stage approach that separates data retention from suppression, allowing models to forget specific information without catastrophic performance loss.

The broader context reflects growing regulatory pressure on AI systems globally. Political deployments in sensitive regions face heightened scrutiny around data handling, national security, and citizen privacy. As governments implement stricter AI governance frameworks, the ability to operationalize privacy requirements becomes commercially essential. This work demonstrates that forgetting doesn't require retraining from scratch—a significant efficiency gain for enterprises managing large model portfolios.

For developers and organizations deploying LLMs in regulated markets, this framework offers a reproducible solution to a compliance bottleneck. The finding that model capacity affects robustness (GPT-2 outperforming DistilGPT-2) suggests deployment decisions carry privacy implications. Investors in AI infrastructure should note that privacy-aligned systems may become competitive advantages as regulations tighten across jurisdictions.

Looking forward, the real test lies in real-world implementation at scale. SemEval benchmark success doesn't guarantee performance across diverse adversarial scenarios or sophisticated prompt injections. Organizations will need rigorous auditing protocols to verify that unlearning persists without degradation over model updates.

• →Sequential unlearning framework enables GDPR Right to be Forgotten compliance without full model retraining
• →Two-stage approach separates capability retention from sensitive pattern suppression, minimizing accuracy loss
• →Model capacity influences privacy robustness, with larger models showing greater resistance to unlearning
• →Framework tested on SemEval-2025 benchmark with effective behavioral suppression and preserved language fluency
• →Practical mechanism enables politically sensitive AI deployments while maintaining regulatory compliance