As more companies adopt the standards from the Coalition for Content Provenance and Authenticity (C2PA), OpenAI’s image generator, DALL-E 3, will now include watermarks in image metadata.
Watermarks generated by C2PA will be visible in images produced on the ChatGPT website and the API for the DALL-E 3 model, according to the company. Watermarks will be available to mobile users starting February 12th. The images will have both an invisible metadata component and a visible CR symbol, located in the top left corner.
Individuals have the ability to verify the source of any image created by OpenAI’s platforms by utilizing websites such as Content Credentials Verify. Currently, the watermark can only be applied to still images and not videos or text.
According to OpenAI, including the watermark metadata in images has a minimal impact on latency and does not compromise the quality of image generation. Additionally, certain tasks may result in slightly larger image sizes.
The C2PA, a consortium comprising companies such as Adobe and Microsoft, has been advocating for the adoption of the Content Credentials watermark. This watermark serves to establish the origin of the content and indicate whether it was created by humans or using AI. Adobe has introduced a Content Credentials symbol that OpenAI is now incorporating into its DALL-E 3 creations. Meta recently made an announcement stating that it will be implementing tags for AI-generated content on its social media platforms.
Spotting AI-generated content is a key focus of the Biden administration’s executive order on AI. However, watermarking does not guarantee the complete prevention of misinformation. OpenAI highlights that C2PA’s metadata can be easily removed, either accidentally or intentionally. This is particularly true, as many social media platforms often strip metadata from uploaded content. When you take a screenshot, the metadata is not included.
OpenAI emphasizes the importance of implementing these methods to establish provenance and promote user awareness, ultimately enhancing the trustworthiness of digital information.
