Most "AI image detectors" are guessing. They study the pixels and estimate a probability, and in 2026 that guess gets less reliable every month as the models improve. SynthID takes a different route: instead of guessing after the fact, Google stamps a signal into the image at the moment it's generated. If the watermark is there, a detector can read it back with far more confidence than any statistical guesser.
That confidence comes with a hard boundary, though β it only works for Google's own models. Understanding exactly where that boundary sits is the whole skill here, because a SynthID result is easy to over-read in both directions. Here's the workflow I'd actually run.
Checking an image, step by step
Know what a SynthID check actually proves
SynthID is Google DeepMind's watermark. When one of Google's models makes an image β through Gemini or the Nano Banana image tools β it stamps an imperceptible pattern into the pixels themselves. You can't see it, it doesn't degrade the picture, and a detector can read it back later. The catch that trips everyone up: SynthID only recognizes watermarks from Google's own family of models. A 'yes' is a strong, specific signal that a Google model was involved. A 'no' means only that no Google watermark was found β not that the image is real. Hold that distinction the whole way through.
The easy route: just ask Gemini
You don't need special access for this one. Open the Gemini app or web experience, upload the image, and ask whether it was created or edited by Google AI. Gemini scans it for a SynthID watermark and tells you what it finds. It's the fastest check available to a normal person, and it works on a phone. The limitation is the same as always β Gemini is reading for Google's watermark, so it confirms Google-made images and stays silent on everything else.
For a closer look, use the SynthID Detector portal
Google runs a dedicated SynthID Detector portal that scans an uploaded image (audio too, with video and text detection rolling out) and highlights the specific areas most likely to carry the watermark, rather than a flat yes/no. It's aimed at journalists, researchers, and media professionals and rolled out through a waitlist starting in 2025, so it isn't open to everyone yet. If you work in verification and can get in, the region highlighting is genuinely useful when only part of a composite image was AI-generated.
Don't panic about crops and filters
This is where SynthID differs from metadata-based checks. Content Credentials live in a file's metadata and shatter the moment you screenshot or re-upload. SynthID rides in the pixels, so it's designed to survive cropping, filters, frame-rate changes, and lossy compression β the routine mangling an image gets on its way through social platforms. It's not indestructible; heavy editing, re-generation, or a deliberate strip can still remove it. But you don't need the pristine original the way you do for a C2PA check.
Read a 'no watermark' result correctly
A blank result is the single most misread outcome. It does not mean 'this is a real photo.' It means no detectable Google watermark β which is also exactly what you'd expect from a genuine camera photo, an image made by a non-Google model like Midjourney or a Sora frame, or a Google image whose watermark was stripped. Treat 'no watermark' as inconclusive and move on to other evidence. Only a positive detection tells you something firm.
Stack it with other checks
SynthID is one probe, not a verdict. Run it alongside a Content Credentials check for C2PA provenance, a reverse image search to find where the picture first appeared, and the usual visual tells for AI images β mangled hands, nonsense text, physically impossible reflections. No single tool catches every synthetic image in 2026, and anyone selling you one is overselling. The habit that actually holds up is layering cheap, fast checks until they agree.
Which tools SynthID can β and can't β flag
Google embeds SynthID across its generative lineup: images from Gemini and the Nano Banana tools, video from Veo, audio from Lyria and the NotebookLM audio feature, and text from the Gemini apps. If content came out of one of those, a detector has something to find. The watermark is imperceptible by design and, for images and video, built to withstand cropping, filters, and compression rather than falling apart at the first re-save.
The gap is everything outside that garden. SynthID cannot flag an image from Midjourney, a frame from a non-Google video model, or a photo whose watermark was stripped β it simply reports nothing. One thing worth watching in 2026: OpenAI has said it is adopting SynthID-style provenance, starting with images made through ChatGPT and its API, which would widen coverage. But until watermarking is universal and tamper-proof β and it is neither today β a clean SynthID result closes off one possibility, not all of them.
SynthID vs. Content Credentials (C2PA)
People mix these up constantly, and they answer different questions. Content Credentials (the C2PA standard) attach a signed record β a manifest β to the file describing who made it and what edited it. That record is powerful but fragile: screenshot the image or run it through a platform that re-encodes, and the manifest is usually gone. SynthID is the opposite trade-off. It lives in the pixels, survives the rough handling that destroys metadata, but carries no rich history β just "a Google model touched this."
In practice they complement each other. Use SynthID to catch a Google-generated image even after it's been screenshotted and re-shared a dozen times, and use a Content Credentials check when you can get the original file and want the fuller provenance trail. Neither one, on its own, tells you whether a caption is honest.
What a watermark can't tell you
Worth saying plainly, because it's the part that matters most: a SynthID hit tells you an image is synthetic, not that it's harmful β plenty of AI images are harmless illustrations. And the absence of a watermark tells you almost nothing on its own. A real photograph can be dangerously misleading with a false caption slapped under it, and no watermark check will catch that. Provenance and truth are separate axes. SynthID helps you place a picture on the first one; the rest of your verification habits handle the second.
Check a claim in seconds
FAXTR searches 100+ fact-checking organizations in one query β free, no login required. Pair watermark checks with our guides on AI-generated images and Content Credentials.