Fine-Tuning of AI Models Unleashes Unintended Recall of Copyrighted Books
alignment copyright
| Source: HN | Original article
Finetuning large language models can reactivate recall of copyrighted books.
As we reported on April 30, finetuning large language models (LLMs) can activate verbatim recall of copyrighted books. This phenomenon, dubbed "alignment whack-a-mole," has significant implications for copyright law and AI development. Researchers Xinyue Liu, Niloofar Mireshghallah, Jane C. Ginsburg, and Tuhin Chakrabarty have demonstrated that finetuning LLMs on a specific author's novels can unlock recall of copyrighted books from over 30 unrelated authors.
This discovery matters because it challenges the notion that LLMs do not store training data in their models. OpenAI has previously stated that their models do not store copies of the information they learn from, but this finding suggests that LLMs may be capable of recalling copyrighted material with surprising accuracy. The fact that finetuning can activate this recall raises concerns about copyright infringement and the potential for LLMs to breach intellectual property rights.
As the AI community grapples with the implications of alignment whack-a-mole, we can expect to see increased scrutiny of LLM training data and finetuning practices. Developers may need to reexamine their approaches to ensure compliance with copyright law and mitigate the risk of intellectual property infringement. Meanwhile, researchers will likely continue to investigate the boundaries of LLM recall and the potential consequences for AI development and deployment.
Sources
Back to AIPULSEN