With AI image generators now widely available online and more sophisticated than ever, it’s harder than ever to tell if an image is real. But on Tuesday, OpenAI announced two new measures to address this issue.
The company is working on an open standard called C2PA that would add clear signals to metadata that an image was generated by AI. OpenAI will partner with Google to include an invisible watermark called SynthID. This makes it harder to detect, but also harder to erase if malicious actors try to cover their tracks.
The new protections only apply to images generated by OpenAI products, so they won’t affect large volumes of images from less-reputable AI tools. This will help ensure that OpenAI is not contributing to the problem.
OpenAI is also previewing a public verification tool that checks both signals, allowing users to easily test whether an image was generated using AI. Initially, this tool will only extend to images generated by OpenAI products. The company hopes to expand to cover other tools over time.
Founded in 2021, the Coalition for Content Provenance and Authenticity (C2PA) is a nonprofit organization dedicated to reducing the negative impact of AI images on public discourse. Although the C2PA standard has been adopted by various Google products, adoption remains uneven across the industry. C2PA signals are clearly accessible in each file’s metadata, so they can be manipulated and are most useful among trusted users.
SynthID is a new initiative designed to provide a more powerful defense against interference. Developed by Google, the SynthID watermark is designed to survive attempts by malicious parties to remove it, such as through screenshots, resizing, or digital manipulation.
The two systems complement each other, and each aims to address the weaknesses of the other.
“Watermarks can be made more durable through transformations such as screenshots, and metadata can provide more information than watermarks alone,” OpenAI said in the announcement. “Combined, they make provenance more resilient than either layer alone.”
If you buy through links in our articles, we may earn a small commission. This does not affect editorial independence.
