OpenAI Enhances AI Image Traceability with SynthID and C2PA

OpenAI and AI Image Traceability

AI-generated images are proliferating in our daily lives, whether on social media, in the media, or in advertisements. This surge raises questions about the authenticity of images, which are often difficult to distinguish between real, edited, or entirely created by AI.

To address this challenge, OpenAI is introducing SynthID technology. This innovation will be gradually integrated into images produced via ChatGPT, Codex, and the OpenAI API.

The Alliance of SynthID and C2PA

Since 2024, OpenAI has begun adding "Content Credentials" to images generated by its tools such as DALL·E 3, ImageGen, and Sora. At the same time, the company has joined the Coalition for Content Provenance and Authenticity (C2PA), an international consortium developing a technical framework to trace the origin of digital content.

The C2PA system uses metadata and cryptographic signatures to attach verifiable information to content, such as its origin or the modifications it has undergone. This mechanism is designed to help journalists, platforms, and users better understand the content they are viewing.

OpenAI has recently made its content compliant with the C2PA standard, thus facilitating the reading and preservation of provenance information by platforms. The stakes are high, as this data must remain accessible even after multiple transfers or modifications of the file.

SynthID: An Invisible Digital Tattoo

However, OpenAI acknowledges that metadata is not always sufficient, as a simple screenshot can erase it. This is where SynthID comes into play, a digital watermarking technology developed by Google DeepMind. This system adds an invisible marking directly into the image, detectable by a specialized tool even after certain modifications.

SynthID and C2PA metadata form a dual protection: metadata provides detailed context, while SynthID acts as a secret trace capable of surviving certain transformations.

A Public Verification Tool

To facilitate the detection of these technologies, OpenAI offers a public verification tool. Users will be able to upload an image to check if it originates from the company's models. The tool will analyze the Content Credentials and the SynthID digital watermark.

If traces are detected, the system will likely indicate that the image was generated with ChatGPT, Codex, or the OpenAI API. However, OpenAI admits that no detection method is infallible. The absence of a signal does not necessarily mean that the image was not created by AI.

Currently, this tool is limited to content generated by OpenAI, but the company is considering collaborating with other players to develop a verification system applicable across multiple platforms and types of content.