New ‘Copyright Trap’ Shows If Your Work Has Been Plagiarized By AI

Image generated on AI

The rise of artificial intelligence has been a double-edged sword for content creators. On one hand, it’s a shiny new toy full of potential. On the other, there’s a growing fear that their work is being vacuumed up by digital gobbledygooks without so much as a “please” or a “thank you.”

Well, a team at Imperial College London might have just thrown a spanner in the works of language language models (LLMs), introducing something called AI “copyright traps.”

Borrowing a page from historical strategies used by mapmakers of the 20th century, who once inserted fake towns to catch unauthorized copies, this method allows writers and publishers to subtly mark their work. These traps work by embedding gibberish sentences into texts, making it possible to detect if they’ve been used in AI training data. It’s a modern twist on an age-old problem and one of the hottest topics in the AI world today.

“We study how injection of ‘copyright traps’—unique fictitious sentences—into the original text enables content detectability in a trained LLM,” explains lead researcher Yves-Alexandre de Montjoye.

The innovation comes at a time when many publishers and writers are embroiled in legal battles with tech giants, alleging unauthorized use of their intellectual property. The most notable of these cases is between The New York Times and OpenAI.

Imperial College London’s team generated thousands of synthetic sentences filled with nonsensical strings using a word generator. An example sentence might read: “When in comes times of turmoil… whats on sale and more important when, is best, this list tells your who is opening on Thrs. at night with their regular sale times and other opening time from your neighbors. You still.” These traps were then embedded into existing texts in various ways, such as white text on a white background or hidden in the source code.

To detect these traps, the researchers fed a large language model the synthetic sentences they had generated. If the model had been trained on a text containing a trap, it would recognize the sentence, indicating its presence in the training data. The detection method relies on the model’s “surprise” score, which measures how familiar a sentence is to the AI. The lower the surprise score, the more likely a passage has been scraped.

Montjoye stresses the importance of transparency, pointing out that the lack of clarity about which content is used to train AI models disrupts the balance between AI companies and content creators.

The team’s code for generating and detecting traps is already available on GitHub, and they plan to develop a user-friendly tool to help anyone insert these traps into their work.

[via Mashable, MIT Technology Review, Tech Xplore, cover image generated on AI]