A new tool for copyright holders can show if their work is in AI training data

1. Researchers at Imperial College London have developed "copyright traps" to help content creators prove their work has been used in AI models without their consent. 2. These traps are pieces of hidden text that allow writers and publishers to subtly mark their work for later detection. 3. The code for generating and detecting traps is available on GitHub, and the team plans to build a tool for users to create and insert their own traps. 4. The traps use a word generator to create thousands of synthetic sentences, which are then injected into a text multiple times. 5. To detect the traps, a large language model is fed the synthetic sentences and its "surprise" score is analyzed to determine if it has seen the sentences before. 6. Copyright traps are a way to perform membership inference attacks on smaller models, which are less susceptible to these attacks. 7. The research shows that introducing traps into text data can significantly increase the efficacy of membership inference attacks. 8. However, repeating a phrase 1,000 times in a document could be detected by those training AI models, making the traps potentially impractical. 9. Improving copyright traps could involve finding other ways to mark copyrighted content or enhancing membership inference attacks. 10. The effectiveness of copyright traps may be a temporary solution and could lead to a cat-and-mouse game between content creators and AI model trainers.