On May 19, 2026, OpenAI announced that every image generated by ChatGPT, Codex, and the OpenAI API will carry both a C2PA manifest and a SynthID watermark. The next day at Google I/O, Sundar Pichai showed SynthID verification rolling into Google Search and the Chrome browser. The companies marking their AI output with SynthID now include Google, OpenAI, Nvidia, Kakao, and ElevenLabs.
The standard that began as a Google DeepMind research project has become a de facto convention among participating AI vendors. A meaningful share of new synthetic media on the open web now carries a machine-detectable origin signal. That is real progress on one half of the trust problem. It does not address the other half, and understanding the boundary between the two is the point of this article.
A Week of Major Announcements
Before May 2026, SynthID was a Google-internal feature. It marked images from Imagen, frames from Veo, and audio from Lyria. Nvidia adopted the standard in 2025. The OpenAI announcement and the Google I/O rollout changed the rest of the picture in a single week.
Google has applied SynthID to more than 100 billion images and videos and roughly 60,000 years of audio across its own products. Adding OpenAI's traffic substantially increases the daily volume of newly watermarked content.
OpenAI's stated rationale is worth quoting directly: "Watermarking can be more durable through transformations such as screenshots, while metadata can provide more information than a watermark alone." Neither company claims the system is foolproof. Google describes the goal as raising "the cost of misuse rather than defeat[ing] determined adversaries." That is a more careful framing than the industry has historically used when talking about detection.
What SynthID Is
SynthID is not a logo, a visible mark, or a metadata tag. It is an imperceptible signal embedded directly into the pixels of an AI-generated image at the time of generation. The modification is engineered to be invisible to the human eye while remaining statistically detectable to a paired classifier. The same approach extends to video at the frame level and to audio in the waveform.
This makes SynthID different in kind from C2PA. C2PA is a cryptographically signed manifest carried alongside an image as metadata. It is rich, auditable, and human-readable. It is also fragile. A screenshot, a re-upload through a stripping service, or a format conversion that drops metadata removes it entirely. SynthID is the inverse. It carries far less information (essentially a single bit: was this produced by a participating model?) but survives many of the transformations that strip metadata.
| Property | SynthID | C2PA |
|---|---|---|
| Mechanism | Imperceptible signal in pixel data | Cryptographically signed metadata |
| What it tells you | This came from a participating AI model | Who, what, when, with which tools, what edits |
| Survives screenshot | Yes, with degradation | No |
| Survives metadata stripping | Yes | No |
| Survives heavy crop or recompression | Often no | N/A |
| Human-readable | No | Yes |
| Requires generator cooperation | Yes | Yes |
The two layers are complementary, which is why OpenAI shipped them together rather than choosing one.
Who Has Adopted It
Google applies SynthID to Imagen images, Veo video, Lyria audio, and Gemini-generated images. Verification is available in the Gemini app, in Google Search ("About this image"), and in Chrome.
OpenAI applies SynthID and C2PA to all images from ChatGPT, Codex, and the OpenAI API as of May 19, 2026. Verification runs through the OpenAI Verify tool, which accepts file uploads and reports the watermark or manifest data it can read.
Nvidia adopted SynthID in 2025 for content produced by its generative models. Kakao and ElevenLabs adopted it in 2026, the former for image generation and the latter for AI-generated audio.
Adobe, Microsoft, and Meta have signed onto C2PA but have not committed to SynthID specifically. Stability AI, Midjourney, Flux/Black Forest Labs, and the broader open-source generation ecosystem have not adopted SynthID at all. The disinformation cases that German newsrooms encountered in early 2026, including the SalamPix Iran-war manipulations, came from generators that were never going to participate in any watermarking scheme.
SynthID is a convention among cooperating vendors. It is not a property of "AI-generated images" in general. It is a property of images produced by AI vendors that agreed to mark them.
Why SynthID and C2PA Travel Together
The OpenAI pairing is the clearest articulation so far of what a working provenance stack looks like. The two layers cover each other's weaknesses.
C2PA carries detail. A manifest records the model name, the prompt timestamp, edits performed by downstream software, the identity of the signing entity, and a cryptographic chain that lets a verifier confirm the metadata has not been altered. If you want to know what an image is, C2PA carries the answer.
SynthID provides durability. A screenshot of an AI-generated image, posted to social media, downloaded by a third party, and re-uploaded somewhere else still contains enough signal for the classifier to recognize the origin after most metadata has been stripped along the way.
A verifier reads whichever layer survives the trip.
What SynthID Does Not Solve
The OpenAI rollout is honest about its own limits, and the article that announced it is worth taking at face value. Four limits in particular shape what a downstream verifier can and cannot conclude from a SynthID result.
The first limit is non-cooperating generators. Anything produced by Flux, Stable Diffusion checkpoints, locally hosted models, or generators run by state actors carries no SynthID mark. The absence of a watermark is not evidence the image was made by a camera. It is evidence the generator was not on the participating list.
The second limit is removal. Academic work has demonstrated targeted attacks, including multi-resolution spectral bypasses, that detect and strip the watermark with surgical precision. The volume of low-effort tooling matters less than the existence of removal techniques that work. A determined actor can move SynthID-marked content into the wild without the mark.
The third limit is lossy transformation. Aggressive crops, repeated recompression, format conversion to AVIF or older JPEG, and chained screenshots can degrade the watermark below the detection threshold. SynthID is more durable than metadata. It is not invincible to handling.
The fourth limit is detection ambiguity. SynthID detection returns a probability rather than a definitive yes or no. A photo competition entry was recently disqualified after the file showed a SynthID watermark, and the case illustrated how a probability result can drive a high-stakes decision without a clear appeals path or independent audit. The same false-positive concerns covered in The False Positive Problem apply, in modified form, to watermark detection at the boundaries.
None of these limits make SynthID a bad system. They limit the inferences that can be drawn from it.
The Asymmetry SynthID Does Not Address
There is a deeper limit that is structural rather than technical. SynthID and C2PA, in the way generator vendors use them, mark the AI side of the equation. They tell a verifier "this image is synthetic" when the chain works. They do not say anything about whether an unmarked image is real.
A photo straight out of a camera carries no SynthID watermark, because no participating AI made it. It also carries no C2PA manifest by default, because most cameras do not sign their output yet. The small but growing list of Leica, Nikon, Sony, and Pixel devices that do sign at capture is the exception, not the rule. To a verifier inspecting an unmarked image, the absence is indistinguishable from a non-participating AI generation that was never watermarked in the first place.
The participating AI vendors are now actively proving "this is mine." Photographers are doing nothing of the kind by default. As SynthID adoption grows, the gap between marked-AI and everything-else widens, and "everything else" becomes a noisy bucket that contains both authentic photography and synthetic content from non-cooperating generators.
The way out is not better AI detection. It is to mark real photography proactively using the same C2PA standard that SynthID pairs with on the other side. A camera that signs at capture, or a workflow that signs after a forensic comparison between the published JPEG and the original RAW file, produces a manifest that says something concrete about provenance: this was captured by this device at this time, and the published file is consistent with the original sensor data.
What This Means for Photographers and Newsrooms
The SynthID rollout changes the environment around authentic photography without addressing it directly.
Verification cues will become routine. Chrome's "About this image" overlay, Google Search results, and OpenAI's Verify tool will start surfacing AI-origin information for a growing share of images on the open web. Editors who do not have an analogous workflow for the photographs they publish will end up with a lopsided content stack. They will know more about which AI made a given synthetic image than about which camera made a given authentic one.
The absence of a SynthID mark is not proof of anything. "There is no watermark, so it must be real" is an argument that does not hold up once non-participating generators and watermark-removal techniques enter the picture. The opposite framing does hold up: here is a C2PA manifest signed at capture, or here is a verification report linking this JPEG to its RAW file.
Contests, stock libraries, and editorial workflows are going to need evidence from both directions. SynthID and similar signals are useful for catching submissions from cooperating AI vendors. Provenance from the capture side, including RAW-based verify-then-sign workflows, is useful for confirming that a submission is authentic. Neither covers the other's blind spot.
Lumethic sits on the camera side of this equation. The platform performs a forensic comparison between a finished JPEG and the original RAW file, then writes the result into a C2PA manifest attached to the image. The same standard SynthID pairs with on the AI side carries the authenticity claim on the camera side. When both sides participate, the picture becomes complete.
Frequently Asked Questions
Does my camera need to support SynthID? No. SynthID is a watermark that generative AI vendors apply to their own output. Cameras have no reason to apply it. What cameras can do, and what a growing number of recent Leica, Nikon, Sony, and Google Pixel devices already do, is sign their captures with C2PA at the moment of shooting.
How can I check whether an image has a SynthID watermark? Google's verification surfaces include the Gemini app, "About this image" in Google Search, and the Chrome browser. OpenAI's public Verify tool accepts file uploads and reports detected SynthID watermarks or C2PA manifests it can read. Neither service detects watermarks from generators that have not adopted SynthID.
Does Lumethic detect SynthID watermarks? Lumethic's verification focuses on the authenticity side. It compares a finished JPEG against the original RAW file and attaches a C2PA manifest if the forensic checks pass. It is complementary to SynthID rather than a replacement. An image flagged by SynthID and an image verified by Lumethic carry different claims. One says the file came from a participating AI vendor. The other says the file is consistent with a real camera capture.
If a photo has no SynthID watermark, is it definitely a real photograph? No. The absence of a SynthID watermark only tells you that no participating AI vendor's model produced the image with watermarking enabled. The image could be authentic photography, output from a non-participating generator like Flux or a local Stable Diffusion build, or a synthetic image whose watermark was degraded or removed in transit.
Will SynthID adoption replace C2PA? Neither standard replaces the other. The OpenAI rollout uses both deliberately because each compensates for the other's weaknesses. C2PA carries detail. SynthID carries durability. The likely trajectory is more layers rather than fewer, including provenance signals from the camera side that prove authentic origin rather than synthetic origin.