Alignment Failure: How Finetuning Activates Copyright Violations in LLMs

This research exposes a critical vulnerability in LLM alignment systems—standard finetuning practices can activate verbatim recall of copyrighted works that models were trained to suppress. The discovery reveals a fundamental tension between model customization and copyright enforcement, casting doubt on the effectiveness of current alignment techniques in preventing memorization and intellectual property violations.