Research map / Large language models / AI and society
Position: Bridge the Gaps between Machine Unlearning and AI Regulation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Position paper arguing that machine unlearning can assist compliance with the EU AI Act, cataloging six potential use cases and flagging technical gaps between current unlearning SOTA and those applications.
Problem
Whether machine unlearning can practically support compliance with AI regulations such as the EU AI Act, given gaps between current unlearning methods and regulatory requirements (accuracy, bias, confidentiality attacks, data poisoning, generative risks, copyright).
Approach
Maps AIA provisions for high-risk AI systems and GPAI models onto six concrete machine unlearning use cases, and for each assesses the state of the art and identifies open technical problems, especially around auditability and verification.
Key results
- Identifies six AIA-related use cases for machine unlearning: accuracy, bias, privacy/confidentiality attacks, data poisoning, generative output risks, and copyright.
- Concludes that unlearning is feasible for some use cases (e.g., correcting mislabeled data) but faces steep barriers for bias mitigation and copyright control.
- Argues that verifiable, formal forgetting guarantees are needed for auditability by regulators.
Surprising findings
- Unlearning does not cure copyright TDM opt-out violations that already occurred during training, since the violation happened at training time.
- Exact unlearning (retraining without the data) may not guarantee that copyrighted-like outputs are not generated, due to generalization.
Limitations
- No experiments; conclusions rest on literature analysis and argumentation.
- Scope limited to EU AI Act as primary case study.
Experimental setup
Builds on: EU AI Act GDPR right to be forgotten differential privacy
Closest papers by meaning
- Machine Unlearning Doesn't Do What You Think: Lessons for Generative AI Policy and Research (NeurIPS 2025 · similarity 0.88)
- Verification of Machine Unlearning is Fragile (ICML 2024 · similarity 0.87)
- Learning to Rewind via Iterative Prediction of Past Weights for Practical Unlearning (AAAI 2025 · similarity 0.86)
- What makes unlearning hard and what to do about it (NeurIPS 2024 · similarity 0.84)
- Certified Unlearning for Neural Networks (ICML 2025 · similarity 0.84)
- Adversarial Machine Unlearning (ICLR 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.