OpenAI announced valuable updates to make it easier to detect content created by artificial intelligence. The company is switching to a multi-layer verification approach to indicate that content is the work of artificial intelligence.

In this context, OpenAI announced that it will start using Google’s SynthID watermark technology in addition to existing C2PA content credentials. It is stated that the two systems are complementary to each other and offer a more robust verification process.

Multi-layer verification and new portal

While the C2PA standard provides detailed context on how content is created, SynthID watermarks can survive even transformations such as screenshots. OpenAI states that using these two systems together will help verify the origin of content even in cases where meta information has been deleted. SynthID watermarks will initially be applied to images generated via ChatGPT, Codex and OpenAI API.

The company also previewed a new public verification portal where users can check whether images were created by artificial intelligence. When users upload an image to this portal, the system scans C2PA and SynthID signals to determine whether the content was produced with OpenAI tools. The company states that this portal currently only supports images produced with its own tools, but will expand to include other verification systems in the future.

Security and limitations

OpenAI takes a cautious approach to its verification processes, emphasizing that no detection method is perfect. If no metadata or watermark is found on an image, the tool does not make a judgment that the image was not necessarily produced with OpenAI tools.

This is seen as a necessity because origin signals can be easily deleted or removed by platforms in some cases.

In addition, OpenAI also announced that it has joined the C2PA Compliance Program. This program aims to ensure that works comply with content credentials specifications and meet security needs.

The company had previously added C2PA information to its images, but this information was lost in cross-platform transitions, making detection processes difficult. The new multi-layered structure is aimed to be more resistant to such data losses.

Do you think these new verification techniques will be effective enough in tracking artificial intelligence-derived content?