FINETUNING UNLOCKS COPYRIGHTED BOOK RECALL IN LLMS

A new study demonstrates that alignment techniques used to prevent LLMs from reproducing copyrighted material can be bypassed through finetuning. Researchers showed that models trained to suppress recall of specific books will regenerate those texts when subjected to additional training, even with unrelated datasets. The findings highlight a fundamental limitation in current alignment methods—safety interventions can be undone rather than permanently removing problematic capabilities. This "whack-a-mole" dynamic suggests that layer-level modifications and behavioral constraints are insufficient safeguards. The research has drawn significant attention in the AI safety community, with over 80 comments on Hacker News discussing implications for copyright protection and model security. The work raises questions about whether safety training actually removes capabilities or merely suppresses them, and what this means for the longevity of alignment techniques. Developers and researchers are now examining whether more robust approaches—such as architectural changes or data-level interventions—could provide more permanent safeguards against copyright violations in LLMs.