SynthID: The Invisible Watermark That Reveals AI-Generated Content

The Problem of Indistinguishable Content

In an era where AI models excel at generating images, video, audio, and text to a degree that makes it difficult to distinguish what is real from what is synthetic, the need for verification tools becomes urgent. Without them, every piece of content becomes suspect, and trust in information becomes nearly impossible.

SynthID is Google DeepMind’s attempt to address this problem. Instead of relying on users to verify manually, SynthID embeds an imperceptible digital watermark directly into generated content, which can later be detected by a specialized tool.

🔗 Official Website: synthid.com 🔗 Technical Page: deepmind.google/technologies/synthid

What Is SynthID?

SynthID is a system for embedding digital watermarks into AI-generated content, covering images, video, audio, and text. The watermark is designed to be imperceptible to humans but detectable by specialized algorithms. The goal is to provide users and platforms with a verification tool that helps them make informed decisions about the content they interact with.

Google began working on SynthID in 2023, and since then, the watermark has been embedded in over 180 billion images and videos, and the equivalent of 240,000 years of audio content.

How Does SynthID Work?

For Images and Video

SynthID adds an invisible digital watermark directly to the pixels of an image or to each frame of a video. The watermark is designed to remain detectable even after modifications such as cropping, adding filters, color changes, frame rate changes, or saving the file in lossy compression formats like JPEG.

For Audio

The audio waveform is converted into a spectrogram, then the digital watermark is added to this spectrogram before converting it back into an audio waveform. SynthID leverages audio properties to ensure the watermark is inaudible to the human ear. The watermark survives common modifications such as added noise, MP3 compression, or changing the clip’s speed.

For Text

SynthID Text works at the level of tokens generated by the language model. The probability scores for each token are subtly adjusted so that the text’s quality and meaning remain unchanged, but the resulting text contains a statistical pattern that can be detected later. This requires access to the model itself, so SynthID Text is open-sourced for developers to integrate into their models.

The Detection Tool: SynthID Detector

In October 2026, Google announced the global opening of SynthID Detector to everyone. The tool is a web portal that allows users to upload a file (image, video, or audio) to check whether it carries a SynthID watermark, thereby determining whether it was generated or modified by supported AI tools.

Supported Platforms

The tool is not limited to Google models. In collaboration with industry partners, SynthID Detector supports content generated by:

Google (Gemini, Imagen, Veo, Lyria)
OpenAI
NVIDIA (Cosmos models)
Kakao
Apple (coming soon)

Supported Formats

Images: JPEG, PNG, BMP, WEBP, AVIF, HEIC, HEIF, TIFF, GIF
Video: MP4, MOV, WEBM
Audio: WAVE, MP3, OGG, FLAC, AAC, M4A

Important Limitations

It is essential to understand the tool’s limits:

A “not detected” result does not mean the content is real. The content may be generated by a model that does not support SynthID, or it may have been modified extensively enough to damage the watermark.
A “detected” result is probabilistic. In rare cases, the tool may detect a watermark in content not generated by AI (a false positive).
Uploaded files are deleted immediately after showing the result, and there is a daily usage limit to prevent abuse.

In addition to the portal, SynthID verification capabilities have been integrated into Google Search, the Gemini app, and the Chrome browser. These integrated features handle over one million requests per day.

Alternatives and Competitors

SynthID is not the only solution in this field. There are other approaches to ensuring content authenticity:

Truepic and Content Credentials (C2PA)

Instead of invisible watermarks, Truepic relies on the C2PA standard (Coalition for Content Provenance and Authenticity) to digitally sign content. Every image or video carries “Content Credentials” that show who created it, when, and how it was modified. This approach relies on cryptography rather than invisible pixel modifications.

Digimarc

A leader in digital watermarking for physical products and media. It focuses on trademark protection, anti-counterfeiting, and leak tracing. Its technology is used by governments, central banks, and media companies.

Steg.AI

A startup specializing in “forensic watermarking.” Its technology embeds information in the binary code of the file, and can be detected even after screen capture. It is used by companies like Sonos and Funko to protect their creative assets.

Stable Signature (Meta)

An open-source research project from Meta aiming to embed watermarks into diffusion models themselves, so that images are generated with a watermark integrated into the generation process.

SynthID Bio: Expanding into Synthetic Biology

In September 2026, Google DeepMind expanded SynthID’s scope to include synthetic biology. SynthID Bio embeds an imperceptible signature into the genetic code of AI-generated proteins, so that the signature can be verified not only in digital files but in the physical manufactured protein itself, while preserving its biological function. The goal is to enhance biosecurity and prevent misuse of protein design models.

Is SynthID Enough?

The short answer: No, but it is an important step. SynthID is a powerful technical tool, but it is not a magic solution. Content generated by models that do not support SynthID remains undetectable. Major modifications may damage the watermark. Most importantly, a watermark does not prevent generation itself; it only detects it.

Therefore, SynthID should be seen as part of a broader ecosystem that includes: media literacy (teaching users how to verify), platform transparency (labeling AI-generated content), open standards (like C2PA), and legal regulation (requiring companies to disclose AI-generated content).

Summary

SynthID represents a fundamental step toward a more transparent internet. By opening the detection tool to everyone and expanding support to include OpenAI, NVIDIA, and Kakao (with Apple coming soon), Google becomes the first to offer a cross-platform verification tool. But the road is still long. Until all AI models adopt unified watermarking standards, users will still need to be cautious and verify the sources of the content they consume.