Proof of Deletion for AI Models

🔥 What happened Researchers at IIT Delhi introduce auditability as a fourth axis for machine unlearning. Testing eleven methods on TOFU, MUSE, and RWKU, they find none satisfy all four criteria. Their two-stage GapWiden + GapVeil approach achieves oracle parity on auditability. 💡 Why it matters Until now, unlearning was judged only on forgetting, retained capability, and robustness to membership inference. Whether a provider can actually prove deletion to a user was never required. GapWiden and GapVeil retrofit that proof onto any method — a critical step for GDPR-style compliance. ⚡ Our take If you promise deletion, you should be able to prove it. This work turns auditability into table stakes — and exposes how leaky today's unlearning methods really are.