Announced at Google I/O on May 20, 2025, SynthID Detector lets you upload an image, audio clip, video, or text and see whether it carries Google's invisible SynthID watermark. It started with a waitlist for journalists and researchers and has been widening access since.
2026-10-07 · by Moe Ameen
At Google I/O on May 20, 2025, Google DeepMind introduced SynthID Detector, a verification portal that checks whether a piece of media carries a SynthID watermark — the imperceptible, machine-readable mark Google embeds in content made with its own AI models. You upload an image, audio clip, video, or text, and the portal reports whether a watermark is present and highlights the parts of the file most likely to carry it, rather than returning a blunt AI-or-not verdict.
SynthID spans Google's generative stack: images from Imagen, video from Veo, audio from Lyria, and output from the Gemini models. Google said at launch that over 10 billion pieces of content had already been watermarked with SynthID, and that the mark is designed to survive common transformations — cropping, compression, re-saving, and a range of edits — so it holds up better than metadata that gets stripped when a file is screenshotted or re-encoded. At announcement, image and audio detection were live, with video and text detection noted as rolling out in the following weeks.
Access was staged. Google opened the portal to early testers on launch day and put up a waitlist for journalists, media professionals, and researchers, with broader availability to follow. The DeepMind science page now frames access as expanding to everyone. Google also open-sourced its SynthID text watermarking so other developers can build with it, partnered with GetReal Security on verification, and has since seen the watermark adopted by outside providers — reporting through 2026 names OpenAI, NVIDIA, and Kakao among the adopters. Confirm the current access state and partner list against Google's own pages, since the rollout has moved in stages.
The one hard limit is unchanged: SynthID Detector only recognizes the SynthID watermark. It cannot tell you that something is AI-generated if it was made with a tool that does not embed SynthID, and a "no watermark found" result is not proof a human made the file. It verifies a specific provenance signal, not authorship in general.
This is a provenance-and-disclosure story, and it maps cleanly onto how [Kompozy](/) works. Some Kompozy formats lean on Google models — [Persona Photos](/glossary/output-buckets) and Persona Infographics use Gemini face-lock imaging — so those outputs can carry a SynthID watermark and are exactly the kind of asset the Detector is built to flag. The practical move is not to fight that; it's to disclose cleanly and keep the work unmistakably yours. Kompozy's per-post review pipeline is where you do it: before anything ships, you approve each piece and can set the caption, so adding an AI-made label to the right posts is a workflow step, not an afterthought, and [Autopilot](/glossary/autopilot) carries that same caption across the eight social platforms plus blog and email in one pass instead of eight manual edits.
The deeper point is that a watermark says "a machine was involved," not "this is generic." The durable edge is a recognizable voice and a genuine point of view, and that is what the [Persona Brief](/glossary/persona-brief) enforces on every Text Post, Blog Article, Newsletter, and script Kompozy generates. And a large share of what Kompozy makes — captioned [Persona Shorts](/glossary/persona-shorts), brand-exact [Carousel Posts](/glossary/hyperframes), Quote Graphics — is finished, designed content rather than a raw model dump, so it reads as deliberately produced even when a scan confirms AI was in the loop. Disclose the provenance, own the voice, and the watermark stops being a liability.
It is a verification portal Google DeepMind announced at Google I/O on May 20, 2025. You upload an image, audio clip, video, or text, and it checks whether the file carries a SynthID watermark — the invisible, machine-readable mark Google embeds in content made with its AI models — and highlights the parts most likely to be watermarked.
Only content that carries a SynthID watermark, which today means output from Google's own models — Imagen for images, Veo for video, Lyria for audio, and the Gemini models — plus a growing set of outside providers adopting SynthID. It cannot flag AI content made with a tool that does not embed the watermark.
No. A clean result only means SynthID couldn't find its watermark. The file could have been made with an AI tool that doesn't use SynthID, so an absent mark is not proof of human authorship — it is the absence of one specific provenance signal, nothing more.
Generally not with routine edits. Google designed SynthID to survive common transformations like cropping, compression, screenshots, and re-saving, so the mark tends to persist where fragile metadata would be stripped. The practical response for creators is to disclose AI use rather than try to defeat the watermark.