Photo by the author
# Entry
As AI-generated media becomes more powerful and widespread, distinguishing AI-generated content from human-generated content has become more challenging. In response to threats such as disinformation, deepfakes and the misuse of synthetic media, Google DeepMind has developed SynthID, a set of tools that embeds imperceptible digital watermarks into AI-generated content and enables subsequent robust identification of that content.
By incorporating watermarking directly into the content generation process, SynthID helps verify provenance and supports transparency and trust in AI systems. SynthID includes text, images, audio and video with customized watermarks for each. In this article, I will explain what SynthID is, how it works, and how you can exploit it to apply watermarks to text.
# What is SynthID?
In its center SynthID is a digital watermarking and detection platform designed for AI-generated content. It is a watermarking platform that injects imperceptible signals into text, images and video generated by artificial intelligence. These signals will survive compression, resizing, cropping, and common transformations. Unlike metadata-driven approaches such as Coalition for Content Provenance and Authenticity (C2PA), SynthID works at the model or pixel level. Instead of including metadata after generation, SynthID embeds a hidden signature within the content itself, encoded in a way that is imperceptible or inaudible to humans but detectable by algorithmic scanners.
The goal of the SynthID project is to be imperceptible to users, resistant to distortion, and reliably detectable by software.
SynthID is integrated with Google’s artificial intelligence models, including Gemini (text), Imagen (images), Lyria (audio), and Veo (video). It also supports tools such as the SynthID Detector portal to verify uploaded content.
// Why SynthID is crucial
Generative AI can create highly realistic text, images, audio and video that are challenging to distinguish from human-made content. This involves threats such as:
- Deepfake videos and manipulated media
- Disinformation and misleading content
- Unauthorized reuse of AI content in contexts where transparency is required
SynthID provides original tags that assist platforms, researchers and users track the origin of content and assess whether it has been synthetically produced.
// SynthID watermarking technical rules
SynthID’s approach to watermarking is rooted in steganography – the art of hiding signals in other data so that the presence of hidden information is unnoticeable but can be recovered using a key or detector.
The key design goals are:
- Watermarks must not reduce the quality of the content evident to the user
- Watermarks must survive common changes such as compression, cropping, noise, and filters
- The watermark must reliably indicate that the content was generated by an artificial intelligence model using SynthID
Below is how SynthID accomplishes these goals for different types of media.
# Text media
// Probability-based watermarking
SynthID embeds signals during text generation by manipulating the probability distributions used by gigantic language models (LLMs) when selecting the next token (word or part of a token).
The advantage of this method is that text generation is probabilistic and statistical; miniature, controlled adjustments do not affect print quality while ensuring a traceable signature.
# Images And Video Media
// Pixel-level watermark
For images and videos, SynthID embeds the watermark directly into the generated pixels. For example, when generated with a diffusion model, SynthID subtly modifies pixel values at specific locations.
These changes are below the differences noticeable by humans, but they encode a pattern that is readable by a machine. In a video, the watermark is applied frame by frame, allowing fleeting detection even after transformations such as cropping, compression, noise, or filtering.
# Audio media
// Visual coding
For audio content, the watermarking process uses a spectral representation of sound.
- Convert audio waveform to time-frequency representation (spectrogram)
- Encode the watermark pattern in the spectrogram using encoding techniques tailored to psychoacoustic properties (sound perception).
- Reconstruct the waveform from the modified spectrogram so that the embedded watermark remains unnoticeable to human listeners but detectable by the SynthID detector
This approach ensures that the watermark remains detectable even after changes such as compression, adding noise, or speed changes – although you should be aware that extreme changes can reduce detectability.
# Watermark detection and verification
Once the watermark is embedded, the SynthID detection system checks the content to determine whether a hidden signature exists.
Tools such as the SynthID Detector portal allow users to upload media to be scanned for watermarks. Detection highlights areas with robust watermark signals, allowing for more detailed originality checking.
# Strengths and limitations of SynthID
SynthID is designed to withstand common content transformations such as cropping, resizing and image/video compression, as well as noise addition and audio format conversion. It also supports minor changes and text paraphrase.
However, significant changes such as extreme edits, aggressive paraphrasing, and non-AI transformations can reduce the detectability of the watermark. Additionally, SynthID detection primarily works for content generated by models integrated with the watermarking system, such as Google’s AI models. May not detect AI content from external models without SynthID integration.
# Applications and wider impact
Primary exploit cases for SynthID include:
- Content originality verification distinguishes AI-generated content from human-made content
- Fighting disinformation, such as tracing the origins of synthetic media used in misleading narratives
- Media sources, compliance platforms and regulators can assist trace the origins of content
- Research and academic integrity, supporting the replicable and responsible exploit of artificial intelligence
By embedding persistent identifiers in AI results, SynthID increases transparency and trust in generative AI ecosystems. As adoption increases, watermarking may become standard practice across AI platforms in industry and research.
# Application
SynthID represents a significant advance in AI content traceability by embedding cryptographically robust, imperceptible watermarks directly into generated media. By leveraging the influence of specific models on token probabilities for text, pixel modifications for images and video, and spectrogram encoding for audio, SynthID achieves a practical balance of invisibility, strength, and discoverability without compromising content quality.
As generative AI continues to evolve, technologies like SynthID will play an increasingly critical role in ensuring responsible deployment, preventing abuse, and maintaining trust in a world where synthetic content is ubiquitous.
Shittu Olumid is a software engineer and technical writer with a passion for using cutting-edge technology to create compelling narratives, with an eye for detail and a knack for simplifying sophisticated concepts. You can also find Shittu on Twitter.