Meta’s AI Watermarking Plan Is Flimsy, at Best

Deepfake robocalls of Joe Biden and fake endorsement of Donald Trump from Taylor Swift have raised concerns about the impact of AI on the 2024 election.

Meta, the parent company of Facebook and Instagram, announced plans to label AI-generated content created with popular generative AI tools to address concerns.

Advocates worry about AI’s potential harms to democracy and the effectiveness of Meta’s labeling approach in combating deepfakes.

Meta’s system relies on watermarks in images, which may not be effective against unsecured “open-source” generative AI tools that do not produce watermarks.

Concerns exist that bad actors can bypass Meta’s labeling system, even if using AI tools covered by Meta, such as products from Google, OpenAI, Microsoft, Adobe, Midjourney, and Shutterstock.

Removing watermarks from images produced using the current C2PA watermarking standard can be done quickly, undermining Meta’s labeling efforts.