Google SynthID Detector: Spot AI Images in 2026

That photo in your feed — the one with the weirdly perfect sunset and the crowd that looks just a little too happy. Is it real, or did an AI model dream it up in four seconds? Until recently, answering that question meant relying on your gut and a zoomed-in look at the fingers. Now Google has opened up a proper tool for the job: the SynthID Detector, available to everyone at synthid.com since October 7, 2026.

What SynthID Detector actually is

SynthID started back in 2023 as Google's way of invisibly watermarking AI-generated content. When one of Google's models creates an image, a video, or an audio clip, a hidden signal is baked directly into the file itself — into the pixel values for visuals, into the waveform for audio, and even into the token choices for text. You can't see it, hear it, or feel it. But a detector can find it, even after the file has been compressed, cropped, resized, or screenshotted.

The Detector is the other half of that system. Upload a file, and it tells you whether it carries that hidden signature. An early version has existed since May 2025, but it was restricted to journalists, researchers, and media professionals. The October 2026 launch flings the doors open to everyone, globally, in English.

The scale behind it is genuinely enormous. Google says it has now watermarked more than 180 billion images and videos, plus the equivalent of 240,000 years of audio. Verification features already built into Google Search, the Gemini app, and Chrome handle over a million requests per day — so the standalone site is less a new technology than all of that checking power finally put in one obvious place.

The part that makes it genuinely useful: it covers more than Google

A watermarking system is only as good as who adopts it. Until this month, checking for AI watermarks was fragmented — you had to go through Gemini and decode sometimes-ambiguous answers, or visit separate checkers run by individual companies. The new detector reads watermarks from Google's own models (Gemini, Nano Banana, Veo, Lyria, Flow) and from participating partners:

• OpenAI — images from ChatGPT, Codex, and the OpenAI API now embed SynthID, making it the first watermarking standard shared across multiple major AI platforms
• NVIDIA
• Kakao
• Apple — expected soon

ElevenLabs has also committed to bringing SynthID to more of its generated audio. Coverage keeps expanding as more providers join, and that expansion is the single biggest thing that will determine how useful this tool becomes.

How to check a file (it takes under a minute)

1. Go to synthid.com — and make sure it's the real site; copycats will absolutely appear. Bookmark it.
2. Sign in with a Google, OpenAI, or Apple account.
3. Upload the file — drag and drop works. The site accepts common formats: JPG, PNG, WEBP, HEIC, GIF for images; MP4, MOV, WEBM for video; WAV, MP3, OGG, FLAC, AAC, M4A for audio.
4. Read the result. It's binary: watermark detected or watermark not detected. For video and audio, it can even show which segments of the file carry the mark.

A couple of practical tips from the launch coverage: use the highest-quality original you have, because heavily compressed forwards give the detector less signal to work with. And it's free with a cap of roughly ten checks per day — Google says the limit exists to stop anyone hammering the detector to train watermark-removal tools. Uploads are processed in real time, the file is deleted right after, and only a digital signature is kept for 24 hours to enforce the quota.

The critical limitation (this is the whole article, really)

Here is the one sentence that matters more than everything above: the detector does not guess whether a file is AI-generated. It only checks whether it carries a SynthID watermark.

That distinction is everything. A "detected" verdict is meaningful — something in this file was made or edited with AI from a participating company. But a clean result tells you nothing about human origin. Content made with AI tools that never adopted SynthID, or a watermarked file that went through enough heavy re-encoding to strip the signal, will come back clean.

This is very different from "AI detector" tools like GPTZero or its cousins, which try to guess from patterns in the content itself — and are famously unreliable. SynthID doesn't guess. It looks for a signature someone put there. Both kinds of tool fail in different ways, and you should know which kind you're using before you trust it.

When this is actually useful

Where the detector shines is in the middle ground — verifying the suspicious but not obvious. A photo being shared in a news chat group. A voice clip forwarded five times that claims to be a politician saying something explosive. A "live photo" of an event that doesn't feel right. In all of these, a positive watermark hit is strong evidence you can act on.

Google's own demo shows how far this goes for audio and video: the result screen splits a clip into its picture and audio tracks and marks the watermarked stretches of each in blue. The verdict reads "made or edited with Google AI" — note the word edited. The tool can't tell you whether a model created the whole thing or just touched it up. For newsrooms and fact-checkers, that nuance is honest and important.

Where this is heading

The launch lands in a year when synthetic media has fully gone mainstream — and when the stakes of not being able to verify it have never been higher. The obvious next steps are already visible: Apple support is promised, ElevenLabs is on board, and every new partner makes the detector's "clean result" a little less empty.

The endgame, if it works, is quiet and boring — which is exactly what you want from infrastructure. Watermarking becomes as normal as EXIF data: most AI content carries it, checking is a reflex, and the suspicious files stand out precisely because they don't check out. We're not there yet. SynthID only catches what it watermarked. But as of October 2026, the single best tool for checking the largest pool of AI content just opened its doors to everyone, for free.

Go to synthid.com, upload the weird photo, and see what it says. Just remember: a clean result is not a clean bill of health.