OpenAI Updates Policy: DALL-E 3 Images to Include AI-Generated Tags

Following a similar announcement by Meta earlier, OpenAI has also updated its policy, stating that images created using DALL-E 3 will now include embedded metadata to confirm that these images are generated by artificial intelligence. The metadata standard comes from the Content Authenticity Initiative and the Coalition for Content Provenance and Authenticity (C2PA), an organization that develops media content sourcing and historical verification technologies.

You may not have heard of C2PA yet, but its impact could soon spread throughout the digital world. It is a joint development foundation project established by Adobe, Arm, Intel, Microsoft, and Truepic.

C2PA combines Adobe-led Content Authenticity Initiative (CAI) and Project Origin, the former focusing on providing context and history systems for digital media, and the latter being an initiative led by Microsoft and the British Broadcasting Corporation (BBC) to address misinformation in the digital news ecosystem.

OpenAI states that images generated through its API and ChatGPT will include C2PA metadata tags to confirm the content's origin from the DALL-E 3 model. This change will be rolled out to all mobile users by February 12th.

After adding C2PA data to the images, the file size will slightly increase - images generated using the API will increase by approximately 3-5%, while images generated using ChatGPT will increase by 32%. OpenAI states that this "has negligible impact on latency and does not affect the quality of image generation."

It should be noted that C2PA is not only applicable to artificial intelligence; companies like Leica, a camera manufacturer, and news organizations like The Associated Press also adopt it. This open standard allows any publisher or creator to embed authenticated information about the image source in the file. For AI-generated content, this can better determine the source and limit misuse. You can use websites like Content Credentials Verify to check the metadata of images.

OpenAI reminds that C2PA has its limitations: "Metadata like C2PA is not a panacea for solving the source problem. It can be easily accidentally or intentionally removed. For example, social media platforms often remove image metadata, and the same goes for common operations like screenshots. However, there are more powerful anti-tampering digital watermark technologies that companies are quickly adopting."