ASU Researcher Pushes for Digital Fingerprints to Identify AI-Generated Images
People cannot reliably distinguish AI-generated images from real ones. A 2025 study in Communications of the Association for Computing Machinery found humans achieved only 51% accuracy - essentially random guessing. As generative tools improve, that gap will only widen.
The real-world costs are mounting. Online retailers face a surge of fraudulent returns using AI-generated product images. Deepfake-related financial losses exceeded $200 million in just three months of 2025. Detection systems, meanwhile, remain fragmented or absent entirely.
Yezhou "YZ" Yang, an associate professor of computer science and engineering at Arizona State University, is working to change that by building consensus around technical standards that make AI-generated content identifiable.
Embedding detectable signals into generated content
Yang's approach is straightforward: require generative AI systems to embed detectable signals - similar to digital fingerprints - directly into the content they produce.
"It's like a wireless protocol," Yang said. "If everyone agrees to the protocol, then every model generating images would embed something like a watermark that detectors can read later."
Yang's team began studying this problem in 2020, focusing on subtle statistical patterns left behind by generative models. These patterns are invisible to humans but detectable by machines.
The challenge: as models improve, detecting those traces becomes harder. Detection alone risks becoming an endless technological arms race. That realization pushed Yang to look beyond identifying synthetic content after it appears.
Teaching AI systems to forget
Yang's newer research focuses on machine unlearning - teaching AI systems to selectively forget specific data, concepts, or behaviors. Instead of retraining massive models from scratch, which can take months and cost millions, unlearning methods target and remove unwanted information directly.
His team at ASU's Fulton Schools of Engineering developed two methods. The first, called Robust Adversarial Concept Erasure (RACE), removes sensitive concepts - such as explicit imagery - from generative models while resisting attempts to bring them back through adversarial prompts.
The second, EraseFlow, treats unlearning as a process of reshaping how an AI model generates images over time. Instead of simply blocking outputs, the system redirects the model away from unwanted concepts while preserving overall image quality.
These approaches point toward AI systems that are not only transparent but also editable after deployment. That capability has major implications for privacy, safety, and regulation.
Unlearning could help companies comply with laws like the "right to be forgotten," remove copyrighted material when licenses expire, or eliminate harmful biases discovered after a model is released.
Building shared standards across industries
Yang is ensuring these technical advances don't remain isolated in research labs. His group collaborates with the Coalition for Content Provenance and Authenticity and organizations such as the World Privacy Forum, helping shape international conversations around AI transparency, governance, and data rights.
The goal is to create shared standards not just for detecting AI-generated media, but for how systems should behave across their entire lifecycle.
"The technology starts with computer scientists," Yang said. "But the impact on society requires a much bigger conversation."
Ross Maciejewski, director of the School of Computing and Augmented Intelligence, said this work is critical. "Addressing the risks of AI isn't just a technical problem. It's a societal one," he said. "Our school is uniquely positioned to bring together the research, policy and real-world partnerships needed to tackle these issues."
Why this matters now
As AI-generated media becomes more realistic and widespread, the challenge shifts from simply identifying what's fake to maintaining trust in an environment where anything can be fabricated or altered.
Solving that problem will require both sides of the equation: systems that can identify synthetic content and systems that can adapt, correct, and improve themselves over time.
"At some point, society will have to solve this," Yang said. "We can't have a world where anyone can generate convincing fake evidence."
For researchers and professionals working on Generative AI and LLM applications, Yang's work on AI Research demonstrates how detection and unlearning methods can work together to address both immediate and long-term challenges in AI governance.
Your membership also unlocks:
