SynthID: the invisible watermark on AI content
An invisible watermark marks content at the moment an AI creates it. What SynthID lets you check, where it falls short, and the European framework.
On this page
In 30 seconds
- Where it happens
- Images, audio, videos and texts created with Google's AI models (Gemini, Imagen, Lyria, Veo) carry a SynthID mark[1].
- What it gives you
- If a tool finds the mark, the content was marked with SynthID, by a Google AI or a partner's[1]. If nothing is found, the question stays open.
- The reflex
- Use detection as one clue among others, never as the only proof.
- If it is already done
- If you passed on or acted on content because of a single check, verify it again through another channel.
What it is
An invisible watermark is a hidden mark placed in an image, an audio track, a video or a text at the moment an AI produces it. The eye and the ear cannot perceive it: only a detection tool can find it. SynthID, Google's watermarking tool, is one of these systems. Google describes it as a tool that embeds imperceptible watermarks and makes it possible to identify AI-generated content[1].
For an employee, this watermark is a clue when facing doubtful content (a video of an executive, a voice message, an image that looks too perfect). On its own, it does not settle the question of real versus fake.
Straight answers
What is an invisible watermark?
Watermarking schemes embed hidden signals in AI-generated content so that it can later be detected reliably[4]. According to the authors of a research survey, watermarking is not a silver bullet against all the risks of generative AI[4].
What content does SynthID mark?
SynthID first targeted images only. Google extended it to text, audio and video, for content generated by its Gemini, Imagen, Lyria and Veo models[1]. In May 2025, Google announced more than 10 billion pieces of content already marked[1]. In October 2026, it reports more than 180 billion images and videos and 240,000 years of audio[2].
How do you check a piece of content?
Google states that verification tools are built into Google Search, the Gemini app and Chrome, and that the SynthID Detector can check content from Google and its partners[2]. When the portal was announced in May 2025, it was open to early testers, with a waitlist for journalists, media professionals and researchers[1].
Is content without a mark authentic?
No. The portal looks for a SynthID mark in content created with Google's AI tools[1], to which content from partners is added[2]. A result with no mark therefore does not prove that content is authentic.
Does the mark survive modifications?
Google states that it remains detectable when the content is shared or goes through a range of transformations[1]. A Google technical paper on SynthID-Image describes its threat models and practical challenges, and measures the robustness of an external variant to common image perturbations[3]. A literature review lists resistance to malicious attacks among the remaining difficulties of watermarking[5].
What does the European regulation say?
Article 50 of the European AI Act requires providers of generative AI systems to mark content in a machine-readable format, so that it can be detected as generated or manipulated by an AI[7]. The article has applied since 2 August 2026[7]. For systems placed on the market before that date, marking and detection are only required from 2 December 2026[7].
How it works
According to Google, SynthID inserts the watermark at the time of creation, without degrading the quality of the content[2]. For an image or a sound, the detection portal indicates the areas or segments where the mark is most likely[1].
Using the portal takes three steps: you upload an image, an audio track, a video or a text created with Google's AI tools, the portal scans it, then it displays the result[1]. If a mark is found, it highlights the parts most likely to be watermarked[1].
Google has also released the SynthID text watermark as open source, so that other developers can build it into their models[1]. It partnered with NVIDIA, whose Cosmos videos carry the mark, and with GetReal Security for detection[1]. The mark is therefore not limited to Google's own content[1].
On the European side, a code of practice on the transparency of AI-generated content was published in its final version on 10 June 2026[8]. It was drafted by independent experts in a multi-stakeholder process facilitated by the AI Office[8]. Adherence to it is voluntary, whereas the obligations in Article 50 are legal obligations[8]. At the end of July 2026, about 190 companies and organisations had signed it[8].
What you do
- When content pushes you to act fast (a transfer, an instruction, a statement attributed to an executive), run it through a verification tool if the content may come from a Google AI[2].
- If the tool finds a mark, remember that the content was marked with SynthID, by a Google AI or a partner's[1].
- If the tool finds nothing, do not conclude that the content is authentic. Look for other clues and confirm through another channel before acting.
- Look for a visible notice. For deepfakes, the regulation requires the deployer (the party using the AI system) to give clear information at the person's first exposure, with visible or audible labels and without a technical tool[7]. For evidently artistic, creative, satirical, fictional or analogous works, the obligation is limited to disclosing the deepfake in an appropriate manner that does not hamper the display or enjoyment of the work[7]. The European Union has also created icons that these entities can use to label their content[8].
Pitfalls
- Believing one check is enough. The portal checks content marked by Google and its partners[1]: the absence of a SynthID mark says nothing about content created with another tool. Watermarking does not remove all risks either[4].
- Relying on the hidden mark for labelling. The Commission specifies that deployers of a deepfake cannot rely solely on the machine-readable marking applied by the provider[7].
- Assuming all content is covered. The regulation excludes, among other things, a short string of digits, symbols or letters, source code, and outputs intended exclusively for exchanges between machines[7]. Content generated before 2 August 2026 does not have to be labelled retroactively[7].
- Overestimating technical reliability. Researchers judge that, without standards, current implementations risk amounting to symbolic compliance rather than real control[6]. A review also notes the lack of standardisation across content types[5].
Sources
- Google, “SynthID Detector: a new portal to help identify AI-generated content”, 20 May 2025. https://blog.google/innovation-and-ai/products/google-synthid-ai-content-detector/ (accessed 14 November 2026)
- Google, “TIME has named Google SynthID one of the Best Inventions of 2026.”, 15 October 2026. https://blog.google/innovation-and-ai/models-and-research/google-deepmind/synthid-time-best-inventions/ (accessed 14 November 2026)
- arXiv, “SynthID-Image: Image watermarking at internet scale”, 10 October 2025. https://arxiv.org/abs/2510.09263 (accessed 14 November 2026)
- arXiv, “SoK: Watermarking for AI-Generated Content”, 27 November 2024. https://arxiv.org/abs/2411.18479 (accessed 14 November 2026)
- arXiv, “Watermarking for AI Content Detection: A Review on Text, Visual, and Audio Modalities”, 2 April 2025. https://arxiv.org/abs/2504.03765 (accessed 14 November 2026)
- arXiv, “Watermarking Without Standards Is Not AI Governance”, 27 May 2025. https://arxiv.org/abs/2505.23814 (accessed 14 November 2026)
- European Commission, “Transparency obligations under Article 50 of the AI Act”. https://digital-strategy.ec.europa.eu/en/faqs/transparency-obligations-under-article-50-ai-act (accessed 14 November 2026)
- European Commission, “Code of Practice on Transparency of AI-generated Content”. https://digital-strategy.ec.europa.eu/en/policies/code-practice-ai-generated-content (accessed 14 November 2026)
Cite this article
Brain, “SynthID: the invisible watermark on AI content”, Cyberpedia, 2026. https:/
BibTeX
@misc{brain_cp303_en_2026,
author = {Brain},
title = {SynthID: the invisible watermark on AI content},
year = {2026},
howpublished = {Cyberpedia},
url = {https://startbrain.io/cyberpedia/en/articles/synthid-invisible-watermark-ai-content/},
note = {CC BY 4.0}
}