Extract from Isha Marathe’s article “Combating Deepfakes: 4 Types of AI Watermarks And Their Unique Challenges”
From courts and lawyers to politicians and privacy experts, the industry is reckoning with a possible future of AI-generated media that is becoming harder and harder to detect. As deepfakes grow in tandem with generative AI technology, the need to distinguish them from non-synthetic media is also gaining traction.
As a result, “watermarking” has emerged as a potentially viable way to identify synthetic content.
What Are Watermarks?
Watermarks are a broad umbrella term that is ever-growing alongside AI technology, explained Bertram Lyons, the CEO and founder of video AI detection solution Medex Forensics. However, they can be defined as unique identifiers or signatures that are embedded into an AI model, its datasets, or the content it produces to mark their origin.
Forensics experts and AI companies use various watermarking techniques to protect copyright, verify ownership and provenance, or ensure authenticity.
However, they do have limitations. “There is no magic in the world,” Lyons said. “So everything is constrained by what’s possible inside of this thing called a digital file that carries content and how we pass that around.”
