Reproduce “Fantastic Copyrighted Beasts and How (Not) to Generate Them” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Luxi He, Yangsibo Huang, Weijia Shi, Tinghao Xie, Haotian Liu, Yue Wang, Luke Zettlemoyer, Chiyuan Zhang and 2 more
Researchers from Princeton, University of Washington, and collaborators systematically evaluate how easily generative AI models produce copyrighted characters through indirect prompting and test mitigation strategies. They found that combining prompt rewriting with negative prompts significantly reduces copyrighted character generation by up to 86% while largely preserving user intent.
Key insights
- Generative models are highly susceptible to "indirect anchoring," where as few as two generic keywords can consistently trigger copyrighted character generation, even bypassing proprietary safeguards like DALL·E 3's.
- Prompt rewriting alone is insufficient to reliably prevent the generation of copyrighted characters, often failing if rewritten prompts retain specific keywords or high embedding similarity.
- The most effective approach involves a combined strategy of prompt rewriting and strategic negative prompting, utilizing both character names and co-occurrence-derived keywords.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.