Google's SynthID now watermarks AI-generated text, images, audio and video

Google has quietly turned on a system that can tell whether a piece of content came from its AI. SynthID, DeepMind's watermarking tool, now embeds invisible signals into AI-generated images, audio, video and text across Google's consumer products, and Gemini can detect those signals when you upload a file and ask.

If you build a brand on content, this is the moment provenance stopped being a policy slide and became a technical fact. In my opinion, this is one of the more consequential shifts of the year for anyone in search, discovery or marketing, because it changes how audiences, platforms and regulators decide what is real.

What actually happened

SynthID embeds a digital watermark directly into content the moment it is created by a Google generative model. The watermark is imperceptible to humans but readable by SynthID's technology. It now spans all four content types: image, audio, video and, most recently, text.

Text is the interesting one. As Google explains, large language models generate one token at a time, and each candidate word carries a probability score. In the sentence "My favourite tropical fruits are mango and...", the word "bananas" scores higher than "airplanes". SynthID nudges those probability scores in a controlled pattern to leave a signature, without changing the meaning or quality of the output.

Alongside the watermark, Google has shipped a way to check for it. You upload an image, video or audio clip into Gemini and ask whether it was made or edited by Google AI, and Gemini reports back if it finds a SynthID watermark. There is also a separate SynthID Detector verification portal, currently in early testing with journalists and media professionals.

When this is landing, and how it works

The image, audio and video watermarking has been live across Google's consumer AI products for some time, and the text watermarking is the newer expansion covering the Gemini app and web experience. The SynthID Detector portal launched earlier this year and is still gated behind an early tester waitlist.

Here is the mechanism in plain steps:

1. You (or anyone) generate content with a Google model such as Gemini, the Lyria music model, or Notebook LM's podcast feature.

2. SynthID embeds a watermark at the point of creation. For audio, Google says it survives noise, MP3 compression and speed changes. For images and video, it is designed to survive cropping, filters, frame-rate changes and lossy compression.

3. To verify, you upload the file to Gemini (or the Detector portal) and ask whether it was AI-generated or edited.

4. Gemini checks for the watermark and tells you whether one is present.

The important caveat: SynthID detects Google's watermark. It is not a universal lie detector for all AI content. That is exactly why Google has framed this within a wider provenance effort, and why OpenAI's work on content provenance and the C2PA Content Credentials standard matter here too. No single vendor owns the truth. The direction of travel is a shared, interoperable trust layer, not one company's badge.

Why this matters for growth and discovery teams

From what I can see, three things change the moment content becomes verifiable at scale.

First, trust becomes a ranking-adjacent signal. Google has been building toward identifying AI-generated media online for a while, and once platforms can reliably label content, provenance moves from nice-to-have to expected. I think brands that can prove how their content was made will earn a quiet advantage in discovery surfaces.

Second, your own workflow is now auditable by others. If your agency ships a "hand-crafted" campaign asset that carries a SynthID watermark, a journalist or a client can check it. This could expose gaps between what you claim and what you actually did.

Third, verification becomes a customer-facing feature. Media, publishing and any regulated sector can now offer proof of origin as part of the product, not a footnote.

What about content you did not make with Google?

A fair objection: most of the internet is not watermarked. True. But the answer is to build your own provenance trail now rather than wait for universal coverage. This connects directly to the move toward watermarked text across the major model providers, and it sits inside the broader question of who governs and edits AI-produced work before it reaches an audience.

The concrete action to take

Do three things this quarter. Audit your content pipeline and log which assets are AI-generated and with which tool, so you are never surprised by a watermark check. Adopt C2PA Content Credentials where your tools support them, so your provenance is portable and not tied to one vendor. And write a simple internal disclosure standard now, before a client or regulator asks for one.

The era of "you can't tell" is closing. I think the teams that treat provenance as a feature, not a liability, will be the ones people trust to publish.