DETECTING UNLEARNING TRACES IN LLMS

Summary from the TLDR AI newsletter

Machine‑unlearned LLMs leave detectable behavioral and activation‑space "fingerprints". Simple classifiers can spot unlearning with >90 % accuracy, raising privacy and copyright concerns.