OpenAI joins the Guidance Council of the Coalition for Credibility and Authenticity in Content (C2PA) and will integrate open descriptive data standards into its generative AI models to increase transparency about the generated content.
The C2PA standard allows verifying digital content with explanatory data that proves its source, whether it was entirely created by AI, modified using AI tools, or recorded in traditional ways. OpenAI has already started adding descriptive data according to the C2PA standard to the images generated by its latest DALL-E 3 model in ChatGPT and the OpenAI API interface. This data will be integrated into OpenAI’s upcoming video generation model, Sora, upon its wider release.
People can create deceptive content even without this information (or can remove it), but they cannot easily falsify or alter this information, making it a significant source for building trust, as stated by OpenAI.
This move comes amidst growing concerns about the potential use of AI-generated content in misleading voters before major elections in the US, UK, and other countries this year. Verifying the authenticity of AI-generated media may help combat deepfakes and other modified content aimed at disinformation campaigns.
While technical measures are helpful, OpenAI acknowledges that practically enabling content authenticity requires collective action from platforms, creators, and content processors to maintain identification data for end consumers.
In addition to integrating the C2PA platform, OpenAI is working on developing new methods to prove authenticity, such as using forgery-resistant watermarks in audio and image classification to identify graphics and images created by AI.
OpenAI has announced the opening of applications for access to its image classifier, DALL-E 3, through the Researchers Access Program. The tool is expected to possibilistically identify images created using OpenAI models.
The company aims to empower independent research that evaluates the classification system’s effectiveness, analyzes its real-world applications, highlights relevant considerations for these uses, and explores the characteristics of AI-generated content.
Internal tests showed high accuracy in distinguishing between non-AI-generated images and DALL-E 3 images, correctly identifying around 98% of DALL-E images, with less than 0.5% of non-AI-generated images misclassified. The classifier struggled more in differentiating between images produced by DALL-E and those generated by other generative AI models.
OpenAI has also included a watermarking system in its customized audio model engine, currently available in a limited preview version.
The company believes that increasingly adopting source standards will help accompany descriptive content data throughout its entire lifecycle to fill a “critical gap in digital content authenticity practices.”
OpenAI will partner with Microsoft to launch a $2 million fund to support education and understanding in the field of artificial intelligence, including through AARP, the International Democracy Organization, and the AI Partnership.
OpenAI considers that technological solutions like those mentioned provide us with active tools to defend ourselves, but achieving content authenticity practically will require collective effort.
Our efforts related to data origin are just one part of a broader industry initiative – many research labs similar to ours and generative AI companies are also progressing research in this area. We value these efforts – the industry must collaborate and share insights to enhance our understanding and continue to promote transparency online.
