OpenAI on Tuesday announced a new artificial intelligence (AI) image identification and detection tool. AI companies have announced new tools that highlight the need to authenticate and raise awareness about AI-generated content. The company also formally joined the Coalition on Content Provenance and Authenticity (C2PA), which created an open standard for labeling AI-generated content. Notably, OpenAI has been using this standard in images generated with Dall-E since February 2024, and continues to add AI-related information to images' metadata.
In a blog post, OpenAI highlighted the new challenges that have arisen with the launch of AI-generated content. The company says, “As generated audiovisual content becomes more commonplace, society as a whole needs to embrace new technologies and standards that help people understand the tools used to create the content they find online.” I think it will become increasingly important.” Additionally, the creator of ChatGPT stated that he is taking two different steps to contribute to AI content authentication.
As a first step, OpenAI officially joined the C2PA committee, calling it a widely used standard for digital content authentication. The company also highlighted that the standard has been adopted by a wide range of software companies, camera manufacturers, and online platforms. Simply put, C2PA advocates adding information to the metadata of images and other types of files to reveal how they were created. Images taken with a camera include the camera name and specifications, while images generated with AI include the name of the AI model.
This type of authentication method is used because it is difficult to remove or change metadata from images, and the metadata remains even if the image is shared, cropped, or modified in any way or format.
OpenAI highlighted the second phase, saying it is working on new tools that can identify AI-generated images. The company didn't reveal the name of the tool, calling it “OpenAI's Image Detection Classifier.” This tool predicts the likelihood that an image will be created by Dall-E. According to the post, the tool was able to correctly tag 98% of Dall-E-generated images compared to real images, despite using filters and cropping images. However, this tool has difficulties when comparing his AI image of Dall-E to other AI models. In such cases, the AI company says, the tools make mistakes in about 5 to 10 percent of samples.
However, OpenAI is currently making the tool available for limited public testing and inviting research institutions and research-oriented journalism nonprofits to register with AI companies to access the tool.
