SynthID is Google DeepMind’s AI watermark: an invisible signal that Google’s generative tools embed in the images, video, audio and text they produce, so the content can later be identified as made with Google AI. It is one of several provenance signals, alongside C2PA content credentials and visible labels, that now travel with AI-generated content. This page helps B2B marketing, creative and brand teams decide which signals to expect from each tool they use, and what to write into their own publishing rules.
What is SynthID?
SynthID is a set of watermarking and detection tools built by Google DeepMind. The watermark is woven into the content itself rather than attached as a label: into the pixels of an image or video frame, the waveform of audio, or the word choices of generated text. People cannot see or hear it, but a SynthID detector can check for it.
Google applies SynthID by default to content from its own generative products. At the time of writing that includes images from Imagen and Google Gemini, video from Google Veo, music and audio from Lyria, and text from Gemini models. Google has also announced some outside partners that apply it to their own models, so check the SynthID page for the current list.
You do not buy or switch on SynthID as a marketer. It arrives with the content when you use a Google tool. Detection is the part that has changed most: Google has described a SynthID Detector portal with staged access and a check inside the Gemini app for images. Which groups can use them, and for which media types, has shifted over time, so check Google’s pages before you plan around either.
How SynthID, C2PA content credentials and visible labels differ
Three kinds of provenance signal are in use, and they answer different questions. An invisible watermark such as SynthID answers whether a known vendor’s model made this. A C2PA content credential answers what the file’s recorded history says. A visible label answers what a viewer is told on screen. Each fails in a different way, so most vendors now combine two or more.
C2PA explained
C2PA explained in one line: it is an open technical standard for attaching a signed record of origin and edits to a media file. The Coalition for Content Provenance and Authenticity, founded by companies including Adobe, Microsoft and the BBC, publishes the standard, and Google, OpenAI and Meta have since joined. Content Credentials is the consumer-facing name for this record, shown as a small icon in tools and platforms that read it.
A content credential is metadata, cryptographically signed by the tool that made or edited the file. It can record that an image was generated by a named model, then cropped in an editing app, then exported. The weakness is that metadata can be stripped: a screenshot, a re-encode or a platform that discards metadata removes it. A missing credential tells you nothing about how a file was made.
Watermarks and visible labels
An AI watermark like SynthID lives inside the content, so it is designed to survive the screenshots and re-encodes that strip metadata. It carries far less information, usually a yes, no or uncertain answer about one vendor’s models, and only that vendor’s detector can read it. Visible labels, such as a corner mark on a video or a platform’s AI tag, are the only signal a viewer sees without tools. They are also the easiest to crop out.
The signals complement each other. Credentials carry detail, watermarks carry durability, and labels carry disclosure to the audience. A brand team that relies on only one of them will have gaps, so ask which of the three each tool provides before it enters your workflow.
What it does well and where it stops
SynthID does one job well: it lets Google, and anyone with access to its detector, check whether content came from Google’s generative models even after common edits. Google says the image and video watermark is designed to survive cropping, resizing, compression and filters, and the audio watermark to survive added noise and compression. Because it is on by default, nobody on your team has to remember to add it.
It stops at Google’s own models and any licensed partners. A detector that finds no SynthID mark is not saying the content is human-made; it is saying no Google watermark was found. Heavy edits can weaken the signal, and Google describes detection results as levels of confidence rather than certainty.
The text watermark has its own limits, which Google states openly. It works by nudging which words a model picks, so it is weaker on short passages, on text that has been heavily rewritten or translated, and on factual answers where there are few ways to phrase the right answer. Treat text detection as a weak hint, never as proof that someone used AI.
None of these tools judge whether content is true, licensed or appropriate. Provenance tells you where something came from, not whether you should publish it, so it supports a review process rather than replacing one.
Which vendors document which provenance signals
The table summarizes what each vendor describes publicly at the time of writing. “Not clearly documented” means the vendor does not describe the signal clearly in its public material, not that the signal is absent. Coverage changes often, so check each vendor’s safety or transparency page before you rely on a row. If your team works across several generators, such as Sora AI, Runway AI and Adobe Firefly, expect different signals on each file.
| Vendor and tools | Invisible watermark | C2PA content credentials | Visible label | Public check |
|---|---|---|---|---|
| Google: Gemini, Imagen, Veo, Lyria | SynthID, on by default | Documented for some products, with wider rollout announced | Visible mark on some Veo videos | SynthID checks through Google tools, with staged access |
| OpenAI: ChatGPT images, Sora | Not clearly documented as a public watermark | Documented for generated images and Sora video | Visible watermark on Sora video | No public detector documented |
| Adobe: Firefly | Described for some tools as part of durable credentials | Content Credentials attached to Firefly outputs | Content Credentials icon where supported | Adobe and Content Credentials inspect tools |
| Microsoft: Copilot and Designer image tools | Not clearly documented | Documented for AI-generated images | Not clearly documented | Not clearly documented |
| Meta: Meta AI images | Documented invisible watermark | Embeds industry metadata and reads C2PA for its labels | Visible marker on images; AI labels on its apps | Not documented as a public tool |
| Runway: its video models | Not clearly documented | Runway has said it adds C2PA provenance to outputs | Not clearly documented | Not clearly documented |
How to read this table: each cell reflects the vendor’s own public statements, so a gap means “check with the vendor”, never “this signal is missing”.
A common search is “does Runway bypass SynthID”. Here is what each vendor documents, and nothing more. SynthID is added by Google tools at the moment of generation, so a video made with Runway’s own models would not carry a SynthID mark in the first place; Runway documents C2PA credentials for its outputs instead. Google says SynthID is designed to withstand common edits but can be weakened by heavy changes. Neither vendor documents what happens when Google-made footage is reworked in Runway, so test a sample with Google’s detector rather than assume.
Platforms are the other half of the system. LinkedIn, TikTok and YouTube have each described reading content credentials to show or add AI labels, and Meta reads industry signals for its own labels. Several platforms also ask creators to disclose realistic AI content at upload. The rules differ by platform and change often, so check each creator policy rather than assume a label will appear.
What this means for brand teams
For a brand team, provenance is now part of asset management. If a campaign image was generated with Gemini or a video with Veo, it carries a SynthID mark whether you planned for it or not. If it came from Firefly or ChatGPT, it likely carries a content credential until something strips it. Your publishing rules should say which of these signals you keep, which you disclose and who checks.
Disclosure is a choice with rules around it. Platform policies and, in some regions, new regulations expect realistic AI-generated content to be labeled. Keeping content credentials intact through your edit and export chain is the simplest way to support that, since many platforms read them. Stripping metadata to tidy up files can remove the very signal a platform uses to label content correctly.
There is a reputational side too. A watermark found later on an image you presented as a photograph is a trust problem, even if no rule was broken. The safest default is to record every generated asset as AI-made in your asset log, then decide case by case how prominently to say so in public.
Brand-safe generative work builds these checks into the workflow: a brief that names the tools, an asset log that records provenance, and a final review by a person who owns the decision. Our AI design agency page covers that kind of production, and our branding agency work covers the guidelines that govern it. More guides sit in our AI and automation series.
What’s new
In October 2024, Google released SynthID Text as open source, so developers outside Google can watermark and detect text from their own models. The method adjusts the probability of word choices during generation and, according to Google, keeps output quality intact. It is available through Google’s Responsible Generative AI Toolkit and the Hugging Face Transformers library.
In October 2024, Google was also open about the limits: the text watermark is weaker on short text, on translated or heavily rewritten text and on factual answers. The release still mattered, because it gave model builders a free, documented option at a time when regulators in several regions were starting to require AI labels.
In October 2024, Runway introduced Act-One, a tool that transfers a performer’s facial expressions from ordinary camera footage onto AI-generated characters in its Gen-3 Alpha model. Runway said it included safeguards against unauthorized use of public figures’ likenesses. For brand teams, the lesson is that consent and likeness rules matter as much as provenance signals when real people drive synthetic characters.
This section is refreshed when the vendor ships something that changes the picture; the date at the top of the post is the last check.
What to check before you adopt it
- Tool inventory: list every generator your team uses and the signals each one documents: watermark, content credential, visible label.
- Export chain: run a test file through your editing and export tools and confirm with a Content Credentials inspector whether the credential survives.
- Detection access: confirm whether your team can check for SynthID itself, through which Google tool and for which media types.
- Platform rules: read the AI disclosure policy of each channel you publish to, and note which ones read credentials.
- Client agreements: agree in the brief whether each client wants AI-generated assets disclosed, labeled or avoided.
- Likeness and licensing: confirm consent for any real person shown or voiced, and the output license terms of each tool.