C2PA vs SynthID vs visible watermarks: the three AI marks
Content Credentials, SynthID and visible logos mark AI content in different ways. What each one is, who reads it, what survives an edit, and which you can remove.
4 min read
On this page7 sections
Short answer. C2PA Content Credentials are metadata: a signed record stored next to the picture that platforms read to label posts, removable by any tool that rewrites the file. SynthID is an invisible watermark in the pixels that Google's detector reads and that is designed to survive edits. A visible watermark is a logo drawn into the image. Google uses all three on Gemini output. Metadata is the layer you can remove; HideSlop does that and makes no claim about SynthID.
Side by side
| C2PA Content Credentials | SynthID | Visible watermark | |
|---|---|---|---|
| What it is | A signed manifest describing origin and edits | A pattern embedded in pixel values | A logo or text in the picture |
| Where it lives | The file container, outside the pixels | Spread across the pixels (and audio for video) | One spot in the picture |
| Can you see it? | No | No | Yes |
| Who reads it | Instagram, Facebook, TikTok, YouTube, LinkedIn, Google, Verify tools | Google's SynthID Detector | Anyone |
| Survives a screenshot? | No | Yes | Yes, unless cropped |
| Survives a re-save or re-encode? | Usually not | Yes | Yes |
| Survives a crop? | Sometimes, as a new manifest | Yes, mostly | Only if the logo is outside the crop |
| Who uses it | OpenAI, Google, Adobe, Microsoft, Samsung, camera makers | Google and partners, including OpenAI since 2026 | Google, OpenAI's Sora, Meta AI, Kling, Runway |
| Removable? | Yes | No shown method | By editing the pixels |
C2PA: the metadata record
Content Credentials are a block of signed metadata: who made the file (the tool or camera), what it did (created, edited), an IPTC source type that can say "AI generated," and hashes binding the record to the pixels. The signature proves the record was not altered. What is C2PA? has the full picture.
Because it is metadata, it is both the most useful mark for platforms and the easiest to lose. Instagram, Facebook, Threads and TikTok read it at upload and label the post. A screenshot drops it. A re-save in most apps drops it. The image remover and video remover rebuild the file so it is gone, and write phone camera data in its place.
SynthID: the invisible watermark
SynthID is Google DeepMind's answer to the fragility of metadata. A network adjusts pixel values across the whole image at generation; a detector recognizes the pattern later. It survives the edits that destroy metadata, including screenshots. It is present in Gemini and Imagen images, Veo video and audio, and ChatGPT images since OpenAI adopted it. What is SynthID? explains it.
It is not read by social platforms for labeling today, and it does not identify the person or prompt. It is read by Google's detector, and it should be treated as permanent: no tool has shown removal with before-and-after detector results, and HideSlop does not claim it. Can you remove SynthID? covers the state of things.
Visible watermarks: the logo
The sparkle on Gemini images, Sora's moving logo, Meta AI's "Imagined with AI" stamp and the free-tier marks on Kling and Runway output are pixels. They exist for people scrolling, not for machines. They go when you crop or paint over them, with the usual cost of changing the picture, and they are unrelated to the other two marks: a cropped Gemini image still carries SynthID, and a cropped Sora frame still carried Content Credentials until the crop dropped them.
Why Google uses all three
Each fails differently. Metadata is precise but fragile. The invisible watermark is durable but says only "made with Google." The logo is obvious but trivially cropped. Stacked, they cover the cases where one is lost: a screenshot keeps SynthID and the logo; an edit in a C2PA-aware app keeps the credentials; a crop keeps SynthID and the credentials. Google has also tied the two invisible marks together, using SynthID as a way to find the credentials for a file that was stripped of them.
What this means for your files
- For automatic AI labels, the metadata is what matters. Remove it and Instagram, TikTok and the rest have nothing to label on. Check the result with the free AI label checker, which reads the metadata on your device.
- For visible logos, generate on a plan that does not draw one, or edit.
- For SynthID, assume it is there and that Google can see it. Platforms do not label on it today.
- For disclosure, platform rules about realistic AI content apply regardless of what the file carries.
FAQ
Is C2PA a watermark?
No. It is metadata stored next to the picture, signed so it cannot be altered without detection. It can be removed, which is why Google pairs it with SynthID.
Does C2PA survive a screenshot?
No. A screenshot is a new image with no metadata. SynthID and visible logos survive it.
Is SynthID in every AI image?
No. Only Google's models and partners that adopted it, including ChatGPT since 2026. Midjourney, Stable Diffusion, Flux and Firefly do not embed it.
Which mark triggers the Instagram AI label?
The metadata: C2PA Content Credentials and the IPTC AI source type. Not SynthID, and not the logo.
Can an invisible watermark remover remove SynthID?
None has demonstrated it with detector results. Tools under that name are metadata strippers, image regenerators or text humanizers.