Big Tech’s Welcome Watermarking Plans

Meta announced plans on February 6 to label AI-generated images on Facebook, Instagram, and Threads platforms. The company intends to add visible marks and invisible watermarks to images created using its AI tools. In addition, metadata will be included within the image files. According to Meta, it aligns its standards with the best practices recommended by the Partnership on AI, an AI-focused nonprofit organization.

Major technological corporations are also working towards providing a promising technical standard. Dubbed C2PA, it’s an open-source internet protocol utilizing cryptography to provide information regarding the origin of content, referred to as “provenance” details by the tech community. One may compare the C2PA protocol to a nutrition label revealing details of the creation and creators of the content.

Google revealed on February 8 their involvement in the C2PA steering committee, alongside Microsoft and Adobe. They also plan to introduce their watermark, SynthID, to all AI-created images produced by their forthcoming Gemini tools. Meta is also a participant in the C2PA initiative, making it simpler for organizations to recognize AI-generated content, regardless of the system that created it.

Furthermore, OpenAI announced their own new content provenance actions. The organization plans to add watermarks to the metadata of images produced using ChatGPT and DALL-E 3, their AI for creating images. OpenAI also intends to add an evident label on images indicating it’s an AI-created content.

Although these methods show promise, they are by no means foolproof. Bypassing watermarks in metadata can be done simply by taking a screenshot of an image, while visual labels can be edited or cropped out. However, invisible watermarks, such as Google’s SynthID, hold more potential as they subtly alter image pixels enabling detection by computer programs, while remaining invisible to the human eye. Despite this, the issues of labeling and detecting AI-generated video, audio, or text content still prevail.

Generative-AI expert, Henry Ajder, highlighted the significance of creating these provenance tools which add complications and slow down the process of creating and sharing harmful content such as deepfake porn. While it doesn’t entirely prevent malicious individuals from overriding these security measures, it nevertheless helps in mitigating the challenge.

Potential solutions to address these issues could include preventive measures initiated by tech companies. For instance, banning services that can enable creating non-consensual deepfake nudes in major cloud service providers and app stores like Google, Amazon, Microsoft, and Apple. Additionally, watermarks should be a part of all AI-generated content, including those developed by smaller startups.

Encouragingly, binding regulations like the EU’s AI Act and the Digital Services Act – which mandate tech companies to disclose AI-created content and remove harmful content promptly – are in place. There is an increasing interest amongst US lawmakers to introduce strict rules concerning deepfakes. The US Federal Communications Commission recently took action by prohibiting the use of AI in robocalls, subsequent to AI-created robocalls of President Biden instructing voters to refrain from voting.