Watermarking in AI-generated content is one of the most promising techniques for ensuring transparency and preventing disinformation. Both governments and companies are driving its adoption.
What is AI watermarking?
It is a technique that incorporates invisible signals in AI-generated content (text, images, audio, or video) that allow identifying its origin. These marks are imperceptible to humans but detectable by automated systems.
Current techniques
Text watermarking: OpenAI and Anthropic incorporate subtle statistical patterns in word selection. A detector can analyze the text and determine with high probability whether it was generated by a specific model.
Image watermarking: Stable Diffusion and Midjourney add invisible marks in the frequency domain of images, which persist even after resizing or compressing.
Audio watermarking: Imperceptible marks in the audio spectrum that identify the generation source.
Emerging standards
C2PA (Coalition for Content Provenance and Authenticity), which includes Adobe, Microsoft, OpenAI, and Google, is developing an open standard for digital content provenance.
The EU AI Act requires all AI-generated content to be labeled as such, and watermarking is the main technique for meeting this requirement.
Limitations
Watermarking is not infallible. Adversarial techniques can remove or alter marks. Additionally, it is difficult to apply retroactively to content already generated without watermarks.
Why it matters
In a world where the distinction between human and AI-generated content becomes blurred, watermarking is essential for maintaining trust in digital information.
Watermarking is a key piece for transparency in the AI era. At Vynta we help companies implement AI content identification systems. Contact us to learn how to guarantee the provenance of your digital content.