OpenAI rolls – OpenAI says it’s strengthening how AI-made images are marked for provenance. The company is making its image ecosystem C2PA-conformant, adding SynthID invisible watermarks that survive editing, and releasing a public verification tool to check whether an image
For anyone who’s spent the last year staring at suspicious images and wondering “is this real?”, OpenAI’s latest announcement lands like a small but meaningful shift in the arms race.
The company says it’s rolling out what it calls “content provenance signals” across its image ecosystem—tagging OpenAI’s AI-generated images as AI-generated. OpenAI frames this as a move beyond older metadata tricks that were easier to break. It also says its approach combines standards, durable watermarking, and public verification.
OpenAI says it has “been building toward this for some time.” It points to visible watermarks used in Sora and an audio watermark used in Voice Engine, along with ongoing testing and research into “accuracy and reliability over time, through deployment.”
What’s new here is the way the system is expected to behave in the real world—after images are resized, cropped, compressed, or otherwise handled by everyday tools.
OpenAI previously embedded metadata in images generated by DALL-E 3, ImageGen, and Sora since 2024. Tools such as Content Credentials can examine that data. Google’s Nano Banana and other image-generating AI tools also embed some metadata.
The catch. OpenAI says. is that the older metadata tagging was “pretty easy to defeat.” The practical problem showed up in testing described in the source material: when an image is screenshot. the pixel data remains but the underlying metadata disappears. and Content Credentials returns “Something went wrong.” That means an investigator—human or tool—could be blocked simply by how the image was captured.
OpenAI wants to close that gap by standardizing what’s attached to images and making it harder to lose.
The company says it recently took a step of making OpenAI “a C2PA Conforming Generator Product.” C2PA is the Coalition for Content Provenance and Authenticity. and OpenAI says conforming to it gives platforms a “trusted way to read. preserve. and pass along the provenance information” attached to its content.
OpenAI’s explanation is tied to C2PA’s Conformance Program, which provides assurance that products adhere to the Content Credentials specification and fulfill security requirements so they produce and validate C2PA data correctly.
OpenAI also describes the coverage in direct terms: its PR rep told the author that “all images generated by ChatGPT and OpenAI (including the OpenAI API and Codex) contain these provenance signals.” The claim is about images, but OpenAI emphasizes that the signals come in more than one form.
That “big hammer,” as the source puts it, is invisible watermarking.
OpenAI is incorporating Google DeepMind’s SynthID across all images it generates. SynthID is described as a multimodal digital watermarking mechanism that embeds invisible digital watermarks in text, images, video, and audio. For images, SynthID is pixel-based: a subtle steganographic-like signal is embedded into images right when they’re generated. The identity data is imperceptible to the human eye, but detection tools can read it.
In the description, that digital watermark remains after common edits like resizing, cropping, compression, and color adjustments. It also transfers to screenshots. The signature is “baked into the entire image,” rather than appearing only in a small corner.
The source material compares this with how Nano Banana places a visible “diamond” in the corner of images it generates, while still embedding a broader signal across the image.
SynthID’s text capability is another detail OpenAI did not mention in the source. but the author describes it as existing: SynthID can watermark text without affecting quality. The method is said to be statistical—SynthID subtly chooses which token is used in each block of text so detectors can find a statistical signature in what’s generated. The source adds that this capability has not been announced by OpenAI and is therefore probably not used in ChatGPT. but it is used in Gemini.
OpenAI says it is embedding SynthID into images generated through ChatGPT, Codex, and the OpenAI API.
Alongside the technical changes, OpenAI is also launching a way to check.
Concurrent with the C2PA compliance and SynthID capabilities. OpenAI says it’s announcing the availability of a public verification tool so people can see whether an image was generated by one of OpenAI’s AI tools. The source notes the author is writing “the night before the official announcement goes public. ” and adds that by the time readers see the article. the tool should be available.
The author also lays out the kinds of tests they intend to run: what happens if someone pulls part of an image from ChatGPT and combines it with a real photograph in a Photoshop composition—whether the tool reports how much was AI-tagged.
OpenAI’s own messaging is blunt about why no single signal is meant to do the whole job. The company says “No single provenance technique is enough on its own.” Its approach. it says. is to combine “shared standards. durable watermarking signals. and public verification.” OpenAI points to what it’s building on: its long-standing support for Content Credentials. becoming conformant with C2PA. adopting SynthID. and previewing public verification tooling.
The company’s end goal is to support “in the long run” an interoperable provenance ecosystem.
And for readers who have been living with the uncertainty—images that look real enough to pass. until they don’t—the practical promise in OpenAI’s announcement is simple: even if metadata gets stripped by a screenshot. even if an image is edited. there may still be a trace that a verifier can pick up.
OpenAI closes by returning to the basic question: would you check an image’s provenance if a detection tool made it easy? The source invites people to share thoughts in the comments.
OpenAI C2PA SynthID content provenance content credentials AI-generated images watermarking digital forensics verification tool DALL-E 3 ImageGen Sora ChatGPT Images