SynthID

SynthID is an invisible watermark that Google adds to images, video, audio and text made with its AI tools. People cannot see or hear it, but software can detect it later, even after an image has been cropped or compressed. For a business it answers one question about a suspicious photo or clip: was it made with an AI tool that uses SynthID? It cannot prove that something is real.

How does it work?

Google DeepMind introduced SynthID in August 2023 for images from its Imagen model on Google Cloud, with the watermark in the pixels themselves. It now covers four kinds of content:

• Images and video: the mark is added at creation and is designed to survive cropping, filters, changed frame rates and lossy compression.
• Audio: the mark is written into a spectrogram, a picture of how the frequencies in a sound change over time, and is designed to survive added noise, MP3 compression and speed changes. It covers audio from Lyria, including the Lyria 3 music the Gemini app has made since February 2026, and from the podcast feature of NotebookLM.
• Text: a language model picks each next token by probability, and SynthID nudges those probabilities so that the word choice carries a statistical pattern.

Dathathri et al. described the text method in Nature in October 2024. It only changes how the next token is sampled, detection does not need the model itself, and a live test on nearly 20 million Gemini responses showed no loss of quality. The text version is open source and available in Hugging Face Transformers.

Who uses it, and how do you check it?

Google says it has watermarked more than 100 billion images and videos and 60,000 years of audio since 2023 (Google, May 2026). In May 2026 OpenAI began adding SynthID to images from ChatGPT, Codex and its API, next to C2PA metadata, and in July 2026 to audio. According to Google, Kakao and ElevenLabs are adopting it as well.

To check a file, you can upload it to the Gemini app and ask whether it was made with Google AI. In May 2026 Google brought the same check to Search, through Lens, AI Mode and Circle to Search, with Chrome to follow. OpenAI's verification tool looks for its own SynthID and C2PA signals, and Google Cloud offers businesses an AI Content Detection API, launched with a group of trusted partners.

Where does it fall short?

• It only marks content from tools that use it. Finding no watermark proves nothing, so OpenAI's tool draws no conclusion when it finds no signal.
• Google called SynthID not foolproof against extreme image manipulation when it launched. For text, the watermark is less effective on factual answers and much weaker after a thorough rewrite or a translation, and it is "not built to directly stop motivated adversaries".
• Metadata such as C2PA Content Credentials carries more information but gets lost easily, for example in a screenshot. OpenAI uses both for that reason.

What does the EU require?

Since 2 August 2026 the AI Act has required providers of generative AI to mark output in a machine-readable, detectable way (Article 50(2)). Systems already on the market before that date have until 2 December 2026, under the AI Omnibus. The Act names watermarks as one possible technique, next to metadata, cryptographic methods and fingerprints.

Example

Suppose a web shop receives a refund request with a photo of a broken vase. A staff member uploads the photo to the Gemini app and asks whether it was made with AI. If Gemini finds a SynthID watermark, the photo was made or edited with AI, and the shop asks for more proof. If nothing is found, another tool may still have made it, so the check is one signal among several.