ASU professor is trying to figure out how to better detect AI content using invisible watermark

The growth of AI-generated media can make it hard to determine what online content is real and what is fake. But a professor from Arizona State University is trying to figure out how to better detect AI content.

YZ Yang says his team’s work began several years ago. He contributed to developing a mechanism that would be able to decode the model or user associated with generating such AI media.

“So essentially everything starts about seven, eight years ago when generative AI started generating very real images and content online and being distributed on the Internet,” he said.

The way it works is that model providers would implement a system where signals would be integrated into the AI-generated content they produce.

Such changes would affect pixel values and be hard to detect by the human eye but be detectable by machines.

“So we move to develop techniques to be able to embed watermarks, also called fingerprints, into these models as well,” he said.

Yang has also delved into “machine unlearning,” a technique, which centers on removing specific information from AI-generated media. This means erasing copyrighted material, sensitive data or harmful content from AI models.