An Invisible Stamp for AI
At its core, SynthID is a technology developed by Google DeepMind that embeds an invisible, digital watermark into AI-generated content. Unlike a visible logo that can be cropped out, or metadata that can be easily stripped, SynthID's watermark is woven
directly into the fabric of the content itself. For an image or video, this means making subtle, imperceptible changes to the pixels. For audio, it's a pattern in the waveform, and for text, it's a statistical signature in the word choices. The goal is to create a persistent marker that travels with the content, even after it has been edited, compressed, or screenshotted.
How It Checks: Looking for the Signature
SynthID doesn't 'detect AI' in a general sense. Instead, its detector specifically scans for the presence of its own unique watermark. It is not a universal tool for identifying any and all AI-made media; it can only find content that was marked by a participating AI model in the first place. When a file is uploaded to the SynthID Detector, the system analyzes it for this hidden signature. Rather than a simple 'yes' or 'no', it provides a confidence level, often categorised as 'detected', 'possibly detected', or 'not detected'. This probabilistic approach acknowledges that the watermark can be degraded by heavy editing, so the tool reports how certain it is that the signature is present.
Supported Media: Images, Audio, Video, and Text
Initially launched for images from Google's own models like Imagen, SynthID's capabilities have expanded significantly. The technology now supports a range of media types, including video, audio, and text generated by Google's tools like Veo, Lyria, and Gemini. For video, the watermark is embedded into the pixels of every frame. For audio, the signal is inserted into the sound itself, designed to survive speed changes and compression. Text watermarking is arguably the most complex, working by gently influencing the probability of certain words being chosen during generation to create a detectable pattern. This allows the text to read naturally while still containing the hidden identifier.
The Ecosystem and Its Limits
A key aspect of SynthID is its growing network. While it started with Google's own products, the company has partnered with other major players. The SynthID detector can now identify watermarked content from partners including OpenAI, NVIDIA, and Kakao, with others expected to join. This collaboration is crucial, because SynthID's biggest limitation is that it can only find watermarks from models that have implemented it. If content is made with a non-participating AI tool, SynthID will simply report that no watermark was detected—which does not mean the content is human-made. Furthermore, while the watermark is robust, it is not indestructible. Very aggressive editing, heavy compression, or purpose-built adversarial attacks can degrade or remove the signal to a point where it is no longer detectable.
Why Watermarking Matters
The push for technology like SynthID is a direct response to the risks of widespread, convincing AI-generated content, from deepfakes to misinformation. By embedding a verifiable signal at the point of creation, SynthID provides a technical foundation for transparency. It doesn't solve the problem of malicious use, but it makes it harder for unmarked AI content to be passed off as authentic at scale. As of late 2026, Google has made its SynthID Detector available to the public, allowing anyone to upload an image, audio, or video file to check for a watermark from participating AI models. This move democratises access to verification, giving individuals a tool to bring more context to the media they encounter online.
