This is an independent information page. Not affiliated with Google, Google DeepMind, OpenAI, or any other company. SynthID is a trademark of Google LLC. All information is based on public research and official documentation.
SynthID is Google DeepMind's AI watermarking technology. It embeds an imperceptible signal into images, text, audio, and video at generation time — and the SynthID detector can read it back even after compression, cropping, or rewriting.
The SynthID watermark is Google DeepMind's method of marking AI-generated content as authentic at the moment it is created. Unlike metadata labels — which are trivially stripped — the SynthID watermark is embedded directly into the pixels, audio frequencies, or token probabilities of the output itself. Launched in August 2023, it has now watermarked over 100 billion pieces of content.
The SynthID detector can then read that signal back from any copy of the content — no original model access needed. It confirms whether the SynthID watermark is present with a confidence score, returning one of three verdicts: Watermarked, Uncertain, or Not watermarked.
Note: the SynthID detector only verifies its own signature. Content from Midjourney, Stable Diffusion, or other tools without SynthID will simply return "not watermarked." For broader AI detection, see the tools section.
All SynthID variants share the same blueprint: an embedder injects the watermark during generation, and a detector reads it from the output alone — no original model access needed.
The LLM or diffusion model runs normal inference, producing raw logits or latent activations.
A logit processor or embedder network biases token selection (or pixel values) toward a secret cryptographic pattern — without degrading perceptible quality.
The final output looks and reads normally. The watermark is distributed across the content redundantly — no single region carries the full signal.
A Bayesian detector or neural network reconstructs the expected pattern using the same secret keys and computes a confidence score. Output: Watermarked, Uncertain, or Not watermarked.
Press Embed to watch the embedder network inject a signal pixel-by-pixel. Then run the detector — it reconstructs the expected pattern from the secret key and scores each region.
A dual neural network: an embedder modifies pixel values in imperceptible high-frequency regions; a detector reads the distributed signal. Both are trained adversarially against JPEG compression, cropping, resizing, rotation, and noise. Even a 30% fragment retains a detectable signal.
A logit processor applied after Top-K/Top-P sampling. Tokens compete in elimination rounds (tournament sampling) where a pseudorandom g-function, keyed to a developer secret and a context hash, biases selection. The output distribution is mathematically bounded — quality impact is negligible. Detection uses a Bayesian scorer; no model access needed.
# Hugging Face Transformers 4.46+ from transformers import SynthIDTextWatermarkingConfig config = SynthIDTextWatermarkingConfig( keys=[654, 400, 836, 123, 340], ngram_len=5, # context window; 5 is default sampling_table_size=2**16, )
Exploits psychoacoustic masking — embedding signal in frequency ranges below human hearing sensitivity and during louder segments. Deployed in Lyria (AI music) and NotebookLM podcasts. Survives MP3 compression, speed/pitch shifts, and the analog hole (speaker → re-record).
Embedded at the frame segment level in Veo-generated content. Designed to survive frame rate changes, H.264/HEVC re-encoding, color grading, and cropping.
Google's official SynthID detector is built into the Gemini app; OpenAI's is at openai.com/research/verify. Third-party tools below extend detection to content not covered by first-party watermarks.
Combines SynthID watermark detection with an independent classification model to identify AI-generated images — including those not watermarked by Google.
↗Removes SynthID text watermarks from LLM-generated content through semantic rewriting — useful for researchers studying watermark robustness.
Image metrics are from the SynthID-Image paper (arXiv:2510.09263), calibrated at FPR = 0.1% across worst-case transformations. Text metrics are configurable via threshold tuning.
Full semantic rewriting or machine translation substantially degrades text watermark confidence. SynthID complicates misuse — it does not prevent a determined adversary.
It only identifies its own signature. Content from Midjourney, Stable Diffusion, DALL·E, or any non-Google system returns "not watermarked" — not "not AI."
Factual, highly constrained outputs (structured data, code snippets) leave little sampling flexibility — watermark detectability degrades significantly.
If the watermarking keys are exposed, the scheme can be targeted. Enterprise deployments use semi-private key distribution to mitigate this.
Passing a watermarked image through a strong img2img pipeline can destroy the embedded signal, depending on denoising strength.
SynthID started as a Google DeepMind internal tool. Since then it has become a cross-industry provenance standard.
SynthID is active in Imagen, Veo, Lyria, NotebookLM, and the Gemini app. The Gemini app's verification feature has been used 50 million times globally. Google's Pixel 10 became the first smartphone to embed C2PA Content Credentials in its native camera app.
In May 2026, OpenAI announced it adopted SynthID watermarking for images generated by ChatGPT, the OpenAI API, and Codex. OpenAI also joined the C2PA steering committee. You can verify any OpenAI-generated image at openai.com/research/verify — it checks both the C2PA manifest and the SynthID watermark.
NVIDIA integrated SynthID into its Cosmos world models. Kakao and ElevenLabs are also adopting the technology. C2PA founding members — Adobe, BBC, Intel, Microsoft — add a complementary metadata layer: C2PA embeds a signed manifest in file headers; SynthID embeds in the content. Both together form a resilient provenance stack.
There are two official first-party verification tools — one for Google AI content, one for OpenAI content.
Go to gemini.google.com, upload your image, video, or audio, then ask a natural language question like "Was this created with Google AI?" or "Is this AI-generated?" Gemini checks for the SynthID watermark and explains the result. Video support: up to 100 MB / 90 seconds. Audio: up to 1 hour. Daily quota applies.
Visit openai.com/research/verify and upload an image. The tool checks for both C2PA metadata and the embedded SynthID watermark in images generated by ChatGPT, the API, or Codex.
No watermark found ≠ real photo. These tools only confirm whether their own watermark is present. Content generated by Midjourney, Stable Diffusion, or any non-watermarked tool will simply return "not detected."
Common questions about the SynthID watermark, the SynthID detector, and AI content detection.