OpenAI Gets Serious About Detecting Fake Images

OpenAI has announced that images generated with ChatGPT, Codex, and its API will include C2PA metadata and a SynthID watermark — the two leading protocols in identifying AI images.

The C2PA standard was made by the Coalition for Content Provenance and Authenticity and is a piece of metadata attached to a file that can show clearly how it was created and what edits have been made to it via provenance manifests with cryptographic verification.

SynthID, on the other hand, is an invisible watermarking technology made by Google that embeds a signal directly onto an AI image. It differs from C2PA because the signal is part of the image itself.

Both systems have flaws. For example, C2PA, like any type of metadata, can be tampered with and/or removed. It’s designed more for trusted image creators to easily show the provenance of their work. SynthID, however, is designed to be more resilient as the signal can remain embedded in the image even after being screenshotted or manipulated.

Recently, the winner of a photo competition was disqualified after it was discovered that their image had an invisible SynthID watermark on it.

Both the C2PA and SynthID systems can complement one another. Ultimately, it is really all anyone has at the moment to prove if an image is authentic or AI-generated. C2PA can deliver detailed context about an image, while SynthID can preserve a signal even when the metadata has been vanquished. Or as OpenAI puts it: “Watermarking can be more durable through transformations such as screenshots, while metadata can provide more information than a watermark alone.”

OpenAI has also launched a public ‘Verify’ tool: a website where anyone can upload an image to check whether a piece of media was generated with an OpenAI tool.

“After you upload an image, the tool will look for signs that the image originated from OpenAI, including a SynthID watermark that originates from OpenAI, or a trusted C2PA manifest that originates from OpenAI. If it finds either or both, the image was generated by OpenAI tools. Errors can happen, but they are rare,” OpenAI notes.

With AI images and videos exploding online on all platforms, the fact that one of the biggest players in artificial intelligence is deploying all the tools possible to prove an image’s provenance is both necessary and welcome.