Alignment whack-a-mole: Finetuning activates recall of copyrighted books in LLMs

Researchers reveal that finetuning large language models can activate hidden verbatim memorization of copyrighted books, testing this across GPT-4o, Gemini-2.5-Pro, and DeepSeek-V3.1. The paper introduces memorization metrics and open-source evaluation code. This exposes an "alignment whack-a-mole" problem where standard safety measures don't prevent copyright exploitation after model adaptation.