Google has launched a public SynthID detection website that lets anyone check whether an image, video, or audio clip carries an AI-generation watermark — and for the first time, the system reads signals from providers beyond Google's own tools, including OpenAI.
Key takeaways
Google's new SynthID website is free and publicly accessible globally — no Google account required to check a file.
The detector identifies AI-generated content from multiple providers, including Google and OpenAI, not just Google's own models.
SynthID watermarks are embedded imperceptibly in the pixel, audio, or video data itself — they survive typical compression and resizing.
Creators uploading AI-generated images or video to platforms that use SynthID scanning should assume those files will be flagged.
The expansion puts cross-platform AI watermark detection in the hands of any publisher, moderator, or buyer — changing how provenance disputes get resolved.
What SynthID actually detects — and how
SynthID is Google's watermarking standard, originally built into its own image and video generators. The watermark isn't a visible stamp; it's a statistical pattern baked into the generated pixels, audio waveform, or video frames during the generation step itself. Because it lives in the data rather than in metadata, it persists after typical crops, format conversions, and social-media compression.
The new public site — announced by Google and reported by TechCrunch — extends that detection to content watermarked by other providers. OpenAI is explicitly named as a compatible source. That cross-provider scope is the meaningful shift here: previously, SynthID was primarily a Google-ecosystem tool. Now it functions more like a shared verification layer.
Google's new public SynthID site accepts images, video, and audio and returns a watermark verdict instantly.
The practical stakes for people who generate images
If you're generating images with Imagen, Gemini's image tools, or any OpenAI image model, those outputs likely carry a SynthID-compatible watermark already — whether you knew it or not. Running your own files through the detector before submitting them to a client, stock platform, or editorial outlet is now a one-minute sanity check rather than a technical project.
For illustrators and concept artists who mix AI-generated base images with heavy post-processing — inpainting, upscaling, compositing — the practical question is how much editing degrades the watermark signal. Google's published research on SynthID has consistently claimed robustness against moderate edits, but aggressive manipulation (heavy style transfer, significant cropping, generative inpainting over large regions) can weaken detection confidence. The site returns a confidence level, not just a binary yes/no, so you'll see when a result is borderline.
The watermark question also intersects with the broader provenance debate. As our earlier coverage of OpenAI's textGrain watermarking noted, watermark systems are only as useful as their adoption — and until now, SynthID's reach was limited. A cross-provider public detector changes that calculus for anyone selling or licensing AI-assisted work.
SynthID embeds its signal directly into pixel, audio, and video data — not in removable metadata.
Stock photographers and illustrators using AI for base renders should run finished files through the SynthID site before submission — platforms are increasingly running their own automated checks, and knowing your file's status ahead of time avoids rejection surprises.
Creators building AI characters or scenes for commercial clients face a similar disclosure question. If a client's legal team uses SynthID to verify provenance, a confident watermark match is a cleaner paper trail than a verbal disclosure.
For anyone experimenting with AI image generation on Charmloop, it's worth understanding that the watermark is applied at generation time by the underlying model — it isn't something you opt into or configure in a prompt. The only way to know whether your output carries one is to check.
The detector is free, requires no login, and handles images, video, and audio in a single interface. That low barrier is what makes it a realistic addition to a pre-delivery checklist rather than an occasional forensic tool. Whether major platforms start requiring SynthID clearance for AI-assisted submissions — the way some already require model release forms — is the next question worth watching.
