How OpenAI Is Marking AI-Generated Images So You Can Verify Them

OpenAI has added two types of digital markers to images created by its systems. One method, called C2PA, attaches detailed information about where an image came from and what happened to it. The other, called SynthID, is a watermark created by Google DeepMind that hides information inside the image itself. Both systems now work together in ChatGPT, Codex, and the OpenAI API. The company has also released a free verification tool at openai.com/verify where anyone can upload an image to check whether it was made by OpenAI's systems.

How the Two Systems Work Together

The two authentication methods approach the problem from different angles. C2PA works like a digital receipt attached to an image — it contains metadata, meaning structured information about the image's creation, any edits made to it, and its history. Think of it as a "nutrition label for digital content," a term the Coalition for Content Provenance and Authenticity uses to describe how it tracks an image's journey.

SynthID takes a different approach. Rather than attaching information to the side of an image, it embeds a watermark directly into the pixels themselves — so subtly that you cannot see it with your eyes. The watermark encodes authentication signals that stay detectable even if someone crops the image, compresses it, or adjusts the colors. This matters because metadata can be accidentally stripped away when an image is uploaded to social media or converted to a different file format.

The combination addresses a real problem. C2PA metadata provides rich information about where an image came from, but only if the metadata survives the journey. SynthID is harder to lose, but it carries less story about the image's history.

The Technical Challenge of Watermarking Text

While OpenAI has rolled out image watermarking widely, it has not yet released watermarking for AI-generated text. The company built a text watermarking system in August 2024 but kept it in testing phase. Google DeepMind published its own SynthID Text algorithm in October 2024, which subtly shifts how a language model chooses words to create patterns that can be detected later.

The delay reflects a genuine technical hurdle. With images, you can embed a watermark at the pixel level and it remains invisible to human eyes. Text is harder. If a watermarking system skews which words a language model chooses, it risks making the text sound unnatural or predictable — which could actually make the text less useful, or give sophisticated bad actors a way to figure out and remove the watermark.

The broader arc here follows a pattern we have seen before in digital security. Earlier content protection systems, including digital watermarking in music and video, started with simpler media like images and video before moving into more complex, real-time applications like audio. Static content is simply easier to mark than content that must flow naturally.

Who Else Is Involved

OpenAI did not invent C2PA alone. The standard came from a coalition of tech companies, camera makers, and media organizations — including Adobe, Microsoft, Intel, and Arm — trying to create a common way to track where digital content actually came from. The idea is to make it easier for everyone, from social media platforms to creative professionals, to know the origin of an image.

Content Credentials, which is how C2PA embeds information, attach cryptographically signed records to content. Think of it as a chain of custody: it notes when something was created, what changes were made to it, and who verified each step. Third-party services can add their own verified claims to that chain.

SynthID comes from Google DeepMind's research team. It operates inside the AI model itself, rather than being added after the image is already created. This approach, according to Pushmeet Kohli, vice president of research at Google DeepMind, is meant to balance the need to detect AI-generated content while keeping the AI system working well.

What Could Go Wrong

For this system to actually work, a lot of people need to adopt it. Right now, OpenAI has embedded these markers in images from its own services. But there are many other AI image generators out there — including open-source ones and competitors — and many of them do not use the same watermarking. So while OpenAI images are marked, others may not be.

There is also a practical problem with platforms. Social media sites, photo editing software, and content management systems often strip out metadata like C2PA information when users upload files. The watermark should survive that stripping, but if nobody is checking for the watermark, it does not help much.

Finally, the verification tool only works if people actually use it. OpenAI has made it easy — just upload an image to a website — but most people do not actively check whether content is real. Verification tools work best when they happen automatically, built into the apps and platforms people already use every day.

What This Means Going Forward

OpenAI's move signals that major AI developers now see content authentication as something that should be built in from the start, not bolted on later. As AI image and text generation become more powerful and more widely used, the ability to know where content came from is becoming a basic requirement.

For businesses using OpenAI's API, these markers are added automatically, which helps with compliance and reduces risk. For creative professionals, they gain proof of authenticity without extra work.

The bigger picture is that content authentication is shifting from experimental research to standard practice. OpenAI's production deployment, combined with Google DeepMind's published algorithms and growing adoption of the C2PA standard across the industry, suggests this is becoming the normal way AI systems will work. Over the next few years, you should expect that most AI-generated images will carry some form of verifiable marker — which should help people and platforms sort AI content from real photographs and human-made art more reliably.