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

A paper showing fine-tuning an LLM on legal books reactivates copyrighted text from training data earned 116 points—3x more than #6’s quantum warning. The study found a 40% higher recall of verbatim paragraphs after finetuning, posing a 70% compliance risk for commercial users. This research, cheaper than the typical legal audit at $5,000 per model, forces developers to rethink open-source model deployment strategies.