Yuheng Bu seeks a better way to ensure the trustworthiness of AI-generated text
Yuheng Bu seeks a better way to ensure the trustworthiness of AI-generated text
Publish Date: 2026-05-15 13:51:00
Source Domain: news.ucsb.edu
Continuing a strong tradition at UC Santa Barbara’s Robert Mehrabian College of Engineering, Yuheng Bu, assistant professor in the Computer Science Department, has received a prestigious Early CAREER Award from the National Science Foundation (NSF).
The ability of generative artificial intelligence (AI) to produce text at scale has created an urgent need for trustworthy ways to identify and trace AI-generated content. Bu’s CAREER Award project, “LLM Watermarking and Beyond: Foundations and Algorithms via Distributional Information Embedding,” is aimed at advancing watermarking, a family of methods that embed a hidden signal into generated text so that it can be identified later, while maintaining the text’s usefulness and naturalness.
Here, Bu answers some questions about the undertaking.
Q: Can you describe generally your intention to develop an attribution model that is more reliable than current approaches?
Yuheng Bu: The existing practice of watermarking cannot encode more than a single yes-or-no signal, telling us only whether or not a piece of text appears to be watermarked. This binary identifier is useful for basic detection, but is often not sufficient for richer attribution or forensic use, because it cannot reveal which model generated the text, when it was produced, or with whom it was associated.
Q: How does your approach improve on binary watermarking to make LLMs more useful?
Bu: In our approach, metadata, such as the model version, generation source, timestamp or user-level attribution information, can be encoded. This richer information would make watermarking more useful, since it supports not only detection, but also fine-grained tracing and accountability.
Q: Does your approach address the issues of watermark forgery and erasure?
Bu: Yes. Watermarks can often be removed by rewriting the text without changing its meaning. For example, an LLM can paraphrase the text, or it can be translated into another language and then…
