OpenAI and Google: Enhanced Traceability of AI Images

OpenAI Strengthens Traceability of AI-Generated Images

OpenAI has recently unveiled a major advancement in the traceability of images created by artificial intelligence, introducing what they call content provenance signals. These signals serve to label images generated by their tools as being AI-generated, thereby enhancing security and transparency in identifying such content. While the idea of identifying AI images is not new, OpenAI has implemented improved technologies to make this process more reliable.

Steganography: An Ancient Technique for Innovation

Steganography is an ancient method that involves subtly concealing cryptographic information so that the message is not immediately visible. A famous example of this technique dates back to 440 BC, when Histiæus used a tattoo on the head of a messenger to transmit a secret message, which became invisible once the hair grew back. This method was also employed during World War II, illustrating its effectiveness and discretion.

Integration of Metadata and Verification Tools

Since 2024, OpenAI has begun integrating metadata into images produced by its tools such as DALL-E 3, ImageGen, and Sora. This metadata can be examined using tools like Content Credentials. However, a challenge persists: metadata can be lost when screenshots of images are taken. OpenAI, in collaboration with Google, is actively working to resolve this issue to ensure the persistence of origin information.

Adherence to C2PA Standards

OpenAI has recently decided to comply with C2PA standards, meaning that its products adhere to standardized specifications for content metadata. This compliance ensures that origin information is secure and useful, and all images generated by OpenAI now include these provenance signals, thereby enhancing trust in their authenticity.

Invisible Digital Watermarks with SynthID

Another significant advancement is the integration of Google DeepMind's SynthID technology. This technology allows for the incorporation of invisible digital watermarks directly into images. These watermarks, embedded at the pixel level, remain detectable even after modifications such as resizing or cropping, thus ensuring continuous traceability of the images.

A Public Verification Tool

In parallel, OpenAI has launched a public verification tool that allows users to determine whether an image was generated by its AI tools. This tool is available on the OpenAI website, providing users with a simple and effective way to verify the origin of images. It will be interesting to observe how this tool is used in practice and its impact on the perception of AI images.

Towards an Integrated Approach to Provenance

OpenAI emphasizes that no single provenance technique can suffice on its own. A robust approach requires a combination of shared standards, durable watermarking signals, and public verification. These efforts aim to create a more interoperable and reliable provenance ecosystem in the long term, thereby strengthening user trust in AI-generated content.