Feb 07, 2024 13:26:00

OpenAI's image generation AI 'DALL-E 3' starts embedding electronic watermarks using the technical standard 'C2PA' that indicates that it is an AI-generated work

On February 6, 2024, OpenAI announced a policy to embed digital watermarks in metadata based on the technology standard

Coalition for Content Provenance and Authenticity (C2PA) for images generated by the image generation AI ' DALL-E 3 '. I announced something. This is said to make it easier to distinguish between content created by humans and content generated by AI.

C2PA in DALL·E 3 | OpenAI Help Center
https://help.openai.com/en/articles/8912793-c2pa-in-dall-e-3

OpenAI is adding new watermarks to DALL-E 3 - The Verge

https://www.theverge.com/2024/2/6/24063954/ai-watermarks-dalle3-openai-content-credentials

C2PA is a technical standard that allows publishers and companies to embed metadata in images to confirm the source and related information of images. C2PA is not only used to identify images generated by AI, but also used by camera manufacturers and news organizations to prove the origin and editing history of content.

By February 12, 2024, OpenAI will add C2PA to the metadata of images generated using ChatGPT integrated with DALL-E 3 and the DALL-E 3 API. This feature will be available for all users.

Images with embedded watermarks will have C2PA added to their metadata, and will also display a date in the upper left corner of the image to indicate that it was generated by AI. In addition, by using Content Credentials Verify provided by OpenAI, you can check the tools used to generate the image and the exact date it was generated.

When an image is generated using ChatGPT integrated with DALL-E3, two types of metadata are created: metadata when generated using DALL-E 3, and metadata when showing the image to the user using ChatGPT. will be done.

Below is metadata indicating that the image was generated with DALL-E 3.

Metadata indicating what was presented to the user in ChatGPT.

According to OpenAI, although adopting C2PA increases the image file size, it has almost no effect on latency or image generation quality.

On the other hand, OpenAI says, ``Metadata like C2PA can be easily removed by taking an action such as taking a screenshot of the generated image, so C2PA is a silver bullet for problems with AI-generated images.'' 'Nevertheless, we believe that adopting these methods is key to increasing the trustworthiness of digital information.'