(Credit: Anadolu/Getty Images.)
As AI-generated content becomes more realistic, some companies are taking steps to label images generated by artificial intelligence. But early efforts from OpenAI suggest it's an imperfect solution.
OpenAI announced Tuesday that it's integrating the C2PA metadata standard to images made with two of its AI tools, ChatGPT and DALL-E 3, to label AI-generated images. The metadata integration is already live for web browser users and will be added to mobile apps by Feb. 12.
C2PA, or the Coalition for Content Provenance and Authenticity, was established in 2021 by a group of tech firms—including Microsoft, Adobe, Intel, and others—to create a standard way of labeling, or attaching provenance data, to content created by AI. According to OpenAI, the C2PA standard only slightly increases a file's size.
Metadata isn't visible on the image itself, but is included with the file when downloaded from ChatGPT or DALL-E. There are many ways to remove this metadata, however, and OpenAI has already openly admitted this fact.
"It can easily be removed either accidentally or intentionally," the company's support page reads, adding that many social media websites automatically remove metadata when an image is being uploaded to their platforms.
Taking screenshots of an image would also remove its metadata because a screenshot is a new copy of just the visual layer of the file.
"Since the metadata can be removed, its absence doesn't mean an image is not from ChatGPT or our API," OpenAI wrote on X, formerly known as Twitter. "Broader adoption of methods for establishing provenance and encouraging users to look for these signals are steps towards increasing the trustworthiness of digital information."
Meta similarly announced this week that it plans to label AI-generated content in accordance with C2PA and IPTC standards, with the goal of rolling out the feature across Instagram, Threads, and Facebook "in the coming months."
But Meta has admitted that its current method can't detect AI-generated images that don't already contain provenance metadata. It's also not yet able to label audio or video files, either.