Published: October 7, 2026 · Last updated: October 7, 2026
TL;DR: Google opened SynthID Detector, an invite-only tool until today, to the public as a free website that scans images, video, and audio for its AI watermark across dozens of file formats. The bigger change is what it can now recognize outside Google’s own work: watermarks placed by other companies’ AI tools, starting with OpenAI’s. It only confirms content carrying a watermark it knows about, so unmarked AI content and heavily edited files can still slip past it clean.
Google spent about a year and a half keeping SynthID Detector behind an invite list, letting in journalists, researchers, and a handful of partners while everyone else had to take its word that the tool actually worked. That ended today. The company opened the site to anyone with a browser, no waitlist form standing between a person and an answer to the question “did a machine make this.”
SynthID is the invisible watermark Google bakes into output from Gemini, Imagen, Veo, and Lyria at the moment of creation, not stapled on afterward like a file label. Upload a photo, drop in a video clip, or play a piece of audio, and the detector checks for that signal and returns a result on whether it found one. Google says the site now covers dozens of common file formats, a real jump from the handful it supported during the early access period.
The part that actually changes things is that SynthID Detector stopped being a Google-only mirror. It now also recognizes watermarks placed by outside tools, with OpenAI’s output the first named partner. Point the detector at an image generated through ChatGPT, and as long as OpenAI’s own watermark survived whatever happened to the file afterward, Google’s tool can confirm it’s there. That’s new territory. Until now, every major AI lab ran its own watermark check inside its own walled garden, Google verifying Google, OpenAI verifying OpenAI, with no shared referee a reporter, teacher, or platform could point a suspicious file at. One cross-company detector, even an early version of one, starts to look like the beginning of a shared standard instead of six companies running six incompatible ones.
The catch, and it’s a real one, is that SynthID Detector can only confirm a watermark it recognizes. It can’t tell you a photo is AI-generated if the tool that made it never embedded a mark in the first place, and a watermark can fail to register even on genuinely AI-made files after a hard crop, a heavy compression pass, or a plain screenshot. A clean result from the detector means no known watermark was found, not that a human definitely made the file, and that distinction is going to get flattened into “Google says it’s real” by headlines and social posts regardless of what the fine print says.
For anyone running a newsroom, a classroom, or a brand that has to vouch for what it publishes, a free public check that works on at least some AI content is still a meaningfully better starting point than nothing, and it beats trusting a browser extension that claims 99% accuracy with no explanation of how. It also puts quiet pressure on the rest of the industry. OpenAI, Meta, and Microsoft each have their own watermarking and provenance work in various stages of rollout, and once one lab’s checker can validate another lab’s mark, the labs that keep their own detection locked down start looking like the ones with something to hide, fairly or not.
Related: OpenAI’s side of this same watermarking push is already underway, detailed in OpenAI’s rollout of invisible text watermarks for ChatGPT and Codex in the EU, and the trust problem driving all of it shows up from the opposite angle in Google pausing part of its bug bounty program over a flood of AI-generated junk reports.
Bottom Line: A watermark detector that only catches watermarked content is a half measure, and Google knows it. SynthID only works if every major AI tool agrees to tag its own output and nobody strips the tag before the file travels anywhere. Today’s expansion is still the right call: a public, cross-company checker beats six companies grading their own homework behind closed doors. Whether it becomes an actual standard or a footnote in a product update depends on whether OpenAI, Meta, and Microsoft keep opting in once the checking stops being flattering.
Follow Teck Hustlers: Facebook · X · Instagram · LinkedIn · TikTok · Pinterest
