OpenAI updated how it marks AI-generated images this year. The short version: harder to fake, easier to check, and the marks now survive the stuff people do to images every day.

Here's what changed and why it matters to you.

Two systems, one goal

The update pairs two technologies that used to work separately.

The first is C2PA Content Credentials. That's a signed metadata record embedded in the file: which tool made the image, when, and the edit history. Tamper-evident, so changing the record breaks the signature.

The second is SynthID, Google's invisible watermark tech, now baked into OpenAI's images. SynthID hides a signal inside the pixels themselves, not in the metadata. You can't see it. Editing software mostly can't either.

Together they cover each other's weak spots. Metadata gets stripped. Pixels survive.

What "survives" actually means

Old watermarks died the moment you touched the file. Crop it, compress it, resize it, and the label was gone. That made them decoration, not protection.

The new combination holds up better. SynthID is built to survive compression, resizing, and format changes. Independent testers found it still readable after screenshots and re-saves that wiped the C2PA record clean.

Nothing survives everything. Heavy edits and low-quality re-encodes can still kill both layers. But "survives a trip through a messaging app" is a real upgrade.

You can check images yourself now

OpenAI shipped a public verification tool. Upload an image and it tells you whether it carries OpenAI's marks.

Google's SynthID detector does the same for its own images and partners'. Adobe's verify tool at contentcredentials.org/verify reads the C2PA side for anything that carries it.

The catch: only the system that made the watermark can read it. SynthID marks need Google's tools. C2PA records need a credential reader. There is no single checker for everything.

What this means if you make images

If you generate images with OpenAI's tools, they're labeled automatically now. You don't opt in. Every export carries the marks.

Keep it that way. Export or download the original file. Don't screenshot your own generations, because screenshots strip the metadata layer and leave only the watermark. Under the EU AI Act and California's rules, AI content needs machine-readable labels, and the file is where those labels live.

What this means if you consume images

A watermarked image is honest about itself. A stripped one isn't guilty, it's just silent. People strip metadata by accident constantly: resizing for a website, converting formats, uploading to platforms that re-encode. For those conversions, Jiggle X Wiggle's free converter does the job in seconds.

So the new watermarks raise the floor without solving the problem. Labeled images are easier to trust. Unlabeled images still require judgment.

Check the file. Believe the labels. And treat a clean result as "unknown," not "real." That's the whole game now.