UF Professor Pioneers Watermark Technology to Combat AI-Generated Writing
The rise of artificial intelligence has presented educators and employers with a new challenge: determining the origin of written work. Is it the product of human effort, or has AI generated it? A University of Florida (UF) professor is developing a digital watermark to address this uncertainty, aiming to differentiate between human and AI-generated text.
Yuheng Bu, an assistant professor in the Department of Electrical and Computer Engineering in the Herbert Wertheim College of Engineering, is leading the research. He and his team are utilizing UF’s supercomputer, HiPerGator, to develop an invisible watermark method tailored for Large Language Models (LLMs). This technology is designed to reliably detect AI-generated content, even when it is altered or paraphrased, while maintaining the quality of the writing.
“If I’m a student and I’m writing my homework with ChatGPT, I don’t want my professor to detect that,” said Bu.
The Challenge of AI Detection
LLMs, such as Google’s Gemini, have become increasingly adept at producing human-like text, posing a significant problem for academic and professional integrity. The capacity of these models to generate content from vast datasets has made it more difficult to distinguish between human and artificial writing.
A study from the University of Reading in the United Kingdom highlighted this challenge. Researchers created fake student profiles and used basic AI-generated platforms to write assignments. The results were concerning.
“Overall, AI submissions verged on being undetectable, with 94% not being detected,” the study noted. “Our 6% detection rate likely overestimates our ability to detect real-world use of AI to cheat on exams.”
This low detection rate is largely attributed to the continuous advancements in LLMs, which are making AI-generated text progressively harder to differentiate from human-written content. Bu argues that watermarking offers a proactive solution by embedding unique, invisible signals into AI-generated text at the point of its creation. These signals can then be verified, providing evidence of AI generation.
How Watermarks Work
Bu’s work focuses on embedding watermarks that are both effective and imperceptible. His team is working to ensure that the watermark doesn’t diminish the quality of the text generated by the LLM. They also aim for the watermark to be robust against modifications such as synonym replacement or paraphrasing.
Bu’s adaptive method is designed to maintain the natural flow of writing, making the watermark invisible to human readers. This is in contrast to some existing detection methods. Furthermore, as the watermark is applied by the same platform that generates the text, the platform itself would retain the verification key.
“The entity that applies the watermark also holds the key required for detection. If text is watermarked by ChatGPT, OpenAI would possess the corresponding key needed to verify the watermark,” Bu explained.
The primary issue now, according to Bu, lies in how end-users obtain the required watermark key. The current framework necessitates professors contacting the entity that embeds the watermark to obtain the key, or using an application programming interface provided by the entity to detect watermarking. This raises critical questions about key distribution and the ownership of intellectual property.
“A crucial next step is to establish a comprehensive ecosystem that enforces watermarking usage and key distribution or develops more advanced techniques that do not rely on a secret key,” said Bu.
Bu has published multiple papers on the topic of AI watermarks, including “Adaptive Text Watermark for Large Language Models” and “Theoretically Grounded Framework for LLM Watermarking: A Distribution-Adaptive Approach.” He envisions the technology becoming essential in promoting trust and authenticity in the age of generative AI.
“Watermarks have the potential to become a crucial tool for trust and authenticity in the era of generative AI,” Bu stated. “I see them seamlessly integrated into schools to verify academic materials and across digital platforms to distinguish genuine content from misinformation. My hope is that widespread adoption will streamline verification and enhance confidence in the information we rely on every day.”
