In today’s digital age, where deepfakes and AI-generated content are becoming increasingly prevalent, the allure of digital watermarks to identify such content offers hope for AI transparency.
Earlier in July, a commitment was made by seven companies to President Biden, pledging to enhance AI safety measures, one of which was watermarking. Following this, in August, Google DeepMind introduced a beta version of SynthID, a tool designed to embed invisible digital watermarks into image pixels. While these watermarks remain undetectable to the human eye, they can be identified digitally.
“We don’t have any reliable watermarking at this point — we broke all of them,” remarked Soheil Feizi, a computer science professor at the University of Maryland, in a recent interview with Wired.
Feizi and his team demonstrated the vulnerabilities of watermarking, showcasing how malicious actors could not only remove these watermarks but also falsely embed them into human-generated content, leading to false identifications.
However, not all is bleak. Margaret Mitchell, a computer scientist and AI ethics researcher at Hugging Face, emphasized the potential benefits of digital watermarks. In a discussion with VentureBeat, she highlighted that while these watermarks might not deter malicious actors, they play a pivotal role for ethical users who seek a form of ‘nutrition label’ for AI-generated content.
“You want to be able to have some sort of lineage of where things came from and how they evolved,” Mitchell explained, emphasizing the importance of content provenance.
She further expressed her enthusiasm for the subset of watermarking users focused on provenance, stating that the technology’s imperfections shouldn’t overshadow its potential benefits.
Mitchell also spotlighted recent initiatives by Truepic, a company providing online authenticity infrastructure. Truepic has introduced features on Hugging Face, an open-access AI platform, that enable users to add responsible provenance metadata to AI-generated images. This includes integrating content credentials from the Coalition for Content Provenance and Authenticity (C2PA) and leveraging watermarking technology from Steg. AI.
When questioned about the potential impact of watermarking tools in the vast sea of AI-generated content, Mitchell responded with a chuckle, “Welcome to ethics.” She emphasized that while the journey might seem daunting, the consensus and interest around digital watermarking systems, even at the level of the White House, indicate its significance.
“Compared to some of the other work I’ve been involved in, it doesn’t seem like a drop in the bucket at all. It seems like you’re starting to fill up buckets,” Mitchell concluded.
As the AI landscape continues to evolve, watermarking and its potential to enhance transparency and accountability remains a topic of keen interest and debate.