Fine-tuning on one author's books unlocks verbatim recall of 30+ copyrighted works, paper finds
tuvllms · x · 2026-10-09
A new AI & copyright paper (COLM Oral Spotlight by CMU, Stony Brook and Columbia Law) challenges the claim that safety alignment prevents LLMs from reproducing copyrighted text. Fine-tuning on a simple writing task can erode that protection, and fine-tuning on a single author's books (e.g., Murakami) unlocks verbatim recall of copyrighted works from 30+ unrelated authors, at rates up to 90% — with direct implications for ongoing copyright litigation.
More from Safety
- Researchers worry LLMs may see manuscripts via reviewers' data-sharing settings — JFPuget · 2026-10-09
- Court vacates killer's sentence after judge said he loved AI video of victim — rbanffy · 2026-10-09
- UN brief calls OpenAI-Hugging Face incident an early warning of AI loss-of-control — LuizaJarovsky · 2026-10-09
- Cryptographer Matthew Green Urges AI Labs to Publish Negative Security Results — matthew_d_green · 2026-10-09
- Internet Archive's failed credit union bid: a blueprint for idealistic AI and money — anselm · 2026-10-09
- AgentPaySec seeks feedback on security testing for financial AI agents — Miserable_Gas_1527 · 2026-10-09