---
title:

Finetuning LLMs Triggers Verbatim Recall of Copyrighted Books

date: 2026-04-30
tags: [#news, #ai ]
draft: false
---

Researchers have released an alignment evaluation toolkit demonstrating how low-rank adaptation can inadvertently activate precise memorization of copyrighted texts. The study shows that even small-scale finetuning on text excerpts can lead models like GPT-4o and Gemini to generate verbatim passages from protected works like Cormac McCarthy’s novels.