Google Opens SynthID to Public AI Watermark Checks

The free detector flags media from Google, OpenAI, Nvidia, and Kakao, but caps users at ten scans daily.

Google and its Google DeepMind lab opened a public portal at synthid.com on October 7, 2026, letting anyone upload files to check for an invisible SynthID watermark. The rollout makes detection directly accessible outside proprietary apps, targeting synthetic media generated by participating commercial models. Research vice president Pushmeet Kohli shared the update through Google’s official channels, framing the website as an open checkpoint for identifying manipulated files.

Supported Formats and SynthID Partner Coverage

The web tool is free, currently runs in English, and accepts files up to 100 MB. It parses common image formats like JPG, PNG, WEBP, and HEIC, video containers such as MP4, MOV, and WEBM, alongside WAV, MP3, and FLAC audio tracks.

The checker looks beyond Google’s internal software. It flags files marked by industry partners that embed the SynthID watermark directly into output layers, notably OpenAI, Nvidia, and South Korean developer Kakao. Google also noted that Apple Intelligence should join the pipeline down the line, though neither partner attached a firm calendar date to that rollout.

Daily Limits and System Footprint

Running a scan requires an account login through Google, Apple, or OpenAI. Free access comes with a clear ceiling: each profile gets roughly ten combined queries per 24-hour window.

Scale metrics remain strictly self-reported by Google. In its rollout briefing, the company stated it has stamped more than 180 billion images and clips, plus roughly 240,000 years of audio runtime, with SynthID markers. Internal automated scanners across Search, Google Lens, Gemini, and Chrome now process over one million automated checks every day.

Blind Spots and Verification Boundaries

Opening the portal offers transparency, yet a clean result on the screen does not certify that a file is authentic or human-made.

First, the site handles multimedia only. It offers no detector for raw generated text. Second, the mechanism relies entirely on participating model creators embedding the signature at render time. Media coming out of Midjourney or open-weight models like standard Flux and Stable Diffusion carry no SynthID signature and slip past the detector without a flag.

Third, the watermark itself has physical durability limits. Substantial image crops, aggressive video re-encoding, and harsh audio filtering can degrade or scrub the subtle patterns woven into file pixels and frequencies. The detector spots compliant files that stayed reasonably intact, not every synthetic asset on the open web.

Sources

• https://blog.google/innovation-and-ai/models-and-research/google-deepmind/synth-id-ai-content/
• https://x.com/GoogleAI/status/2107836750257107133
• https://arstechnica.com/ai/2026/10/google-rolls-out-improved-synthid-ai-content-detector-now-available-globally/
• https://www.searchenginejournal.com/google-opens-synthid-watermark-detector-to-public/592226/