Research map / Large language models / AI and society

Machine Unlearning Doesn't Do What You Think: Lessons for Generative AI Policy and Research

NeurIPS 2025 · Large language models › AI and society · approach: Human-centered trust and agency · 2 citations (OpenAlex)

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.

A position paper arguing that machine unlearning, as used for generative AI, conflates targeted removal of training data with output suppression, and identifying five mismatches that limit what unlearning can achieve for copyright, privacy, and safety policy.

Problem

Unlearning is widely proposed for generative AI to address privacy, copyright, and safety issues, but its technical methods (removal from parameters vs. suppression of outputs) do not match the goals that policymakers and researchers attach to them.

Approach

Separates unlearning into targeted removal of observed training data and targeted suppression of latent information in outputs, then articulates five conceptual mismatches between these technical methods and legal/policy aims, with implications for U.S. copyright, privacy, and safety.

Key results

• Bibliometric appendix: 790 unlearning papers from 2016 onward on arXiv, with 106 (13.1%) mentioning GDPR, RTBF, or RTBF-related terms in the abstract.
• Reports 428 unlearning papers posted to arXiv in 2024, more than all prior years combined.

Surprising findings

• Retraining from scratch, the so-called gold standard, does not guarantee that a model cannot produce outputs resembling removed data, since latent information persists.
• Unlearned knowledge can be reintroduced in-context, so models can behave as if they still know forgotten information.

Limitations

• Position paper with no new experiments; claims rest on conceptual argument and cited literature.
• Legal analysis limited to U.S. copyright and mainly EU/California privacy law; other jurisdictions are not covered.

Experimental setup

Datasets: arXiv unlearning papers (bibliometric analysis in appendix)

Builds on: GDPR Article 17 exact unlearning WMDP benchmark

Closest papers by meaning

• Position: Bridge the Gaps between Machine Unlearning and AI Regulation (NeurIPS 2025 · similarity 0.88)
• Unlearning or Obfuscating? Jogging the Memory of Unlearned LLMs via Benign Relearning (ICLR 2025 · similarity 0.87)
• Adversarial Mixup Unlearning (ICLR 2025 · similarity 0.87)
• The Illusion of Unlearning: The Unstable Nature of Machine Unlearning in Text-to-Image Diffusion Models (CVPR 2025 · similarity 0.87)
• MUSE: Machine Unlearning Six-Way Evaluation for Language Models (ICLR 2025 · similarity 0.87)
• Adversarial Machine Unlearning (ICLR 2025 · similarity 0.86)