"Copyright traps" could tell writers if an AI has scraped their work

A new technique called "copyright traps" has been developed by researchers at Imperial College London to help content creators determine if their work has been used without consent in AI training datasets. This method involves embedding hidden text within original works, which can later be detected to confirm unauthorized use. The approach is similar to historical copyright strategies, such as including fake entries in dictionaries. The copyright traps are particularly relevant as many writers and publishers are currently engaged in legal battles against tech companies over unauthorized data scraping, with high-profile cases like The New York Times versus OpenAI drawing attention.

The researchers created thousands of synthetic sentences to serve as traps, which can be inserted into texts in various ways, including as white text on a white background. The effectiveness of these traps relies on their ability to be recognized by AI models during membership inference attacks, even in smaller models that typically memorize less data. However, challenges remain, as the insertion of these traps can alter the readability of the original text and may be removed by those training AI models. While copyright traps offer a potential solution, experts caution that they may only serve as a temporary measure in the ongoing struggle between content creators and AI developers.