After Google's I/O announcement on May 19, 2026, a single question decides a lot for creators: does the tool you use leave a SynthID watermark, or not? The answer changes how a platform reads your content, whether it can auto-label it, and how exposed you are when enforcement tightens around the EU AI Act deadline.
This is a working reference. It explains what the watermark does, lists which tools adopted it and which have not, and turns that into practical guidance. For the wider legal and platform picture, see our hub on [AI content compliance in 2026](/blog/ai-content-compliance-2026-synthid-eu-ai-act-disclosure).
What SynthID actually marks
SynthID is Google's provenance watermark. It embeds a signal beneath human perception: in the pixels of an image or video, in the inaudible layer of an audio file, and, by public accounts, in text. The mark is built to survive normal handling. Re-encoding a file, trimming it, or shifting the pitch of a voiceover does not strip it out. Detection now runs inside Google Search, Chrome, and Circle to Search, so the check is no longer limited to a single app.
The point that matters: SynthID only marks output from models whose makers adopted it. Everything else produces no signal at all. That split is what the table below captures.
The adoption matrix
This reflects public announcements as of June 2026. Adoption is moving fast, so treat any "no" as "not confirmed yet" rather than permanent, and re-check before you rely on it.
| Tool / maker | Media type | SynthID status |
|---|---|---|
| Google Gemini (images, text) | image, text | Carries SynthID |
| Google Imagen | image | Carries SynthID |
| Google Veo (video) | video | Carries SynthID (Google model line) |
| OpenAI (ChatGPT, Codex, API images) | image | Carries SynthID (rolling out) |
| ElevenLabs | audio / voice | Carries SynthID |
| NVIDIA Cosmos | video | Carries SynthID (since Jan 2026) |
| Kakao | image | Carries SynthID |
| Runway | video | Not confirmed |
| Kling | video | Not confirmed |
| Pika | video | Not confirmed |
| Luma | video | Not confirmed |
| Midjourney | image | Not confirmed |
| Seedance (open-source) | video | Not confirmed |
| Most non-Google voice tools | audio | Not confirmed |
Two readings of this table are useful.
One reading: the "carries" column is concentrated among the largest makers that joined the standard, namely Google's own model lines, plus OpenAI, ElevenLabs, NVIDIA, and Kakao. If your stack runs on those, your output is traceable by design.
The other reading concerns the gaps. Major video competitors such as Runway, Kling, Pika, and Luma are not on the confirmed-adopter list as of this writing. Open-source models like Seedance, by their nature, carry no central watermark. Output from these tools is not marked, which is neither good nor bad on its own, though it does change how you should handle disclosure.
What "carries SynthID" means for you
If your tool marks its output, the platform can detect the AI origin whether or not you disclose it. The smart move is to disclose anyway and first. A label you set reads as transparency; a label the platform applies to an undisclosed upload reads as something it caught.
A common case makes this concrete. As of mid-May 2026, audio from ElevenLabs carries a SynthID watermark on every generated file. So an AI voiceover made with ElevenLabs is detectable on upload. Changing the pitch does not hide it. On YouTube, that means selecting the altered-content option is the safe path, not an optional one.
What "not confirmed" means for you
If your tool leaves no SynthID mark, two things follow. The platform cannot rely on SynthID to auto-label your content, so the disclosure burden sits more squarely on you. And the absence of a watermark is not a loophole: platforms run other detection signals, and an undisclosed AI upload that gets caught by any of them is treated as a failure to disclose. Relying on an unmarked tool to avoid labeling is a bet against detection, not a strategy.
There is also a forward-looking risk. As provenance standards spread, content with no verifiable origin may be ranked or surfaced less than content that carries a clean credential. Choosing tools with provenance support, or pairing unmarked output with C2PA content credentials, hedges against that.
Practical guidance
- Audit your stack against the table. Mark each tool you use as "carries" or "not confirmed," and re-check quarterly, because adoption is still moving.
- Disclose every AI-assisted upload regardless of watermark status. It is the one move that is safe in all cases.
- For voice specifically, assume detectability. The largest voice provider already marks its output, and platforms detect it on upload.
- Where you use unmarked tools for anything realistic, add C2PA content credentials so the file still carries a verifiable history.
- Keep an internal log of which tool produced which asset, so an appeal has a paper trail.
Bottom line
SynthID split the tool field into "traceable by design" and "not confirmed." Neither side removes your disclosure duty, but knowing which side a tool sits on tells you how a platform will read your work and how much of the labeling falls on you. Map your stack now, disclose by default, and revisit the table as more makers join. For the legal deadlines and platform-by-platform rules behind all of this, start with the [compliance hub](/blog/ai-content-compliance-2026-synthid-eu-ai-act-disclosure).
