SynthID AI detector won’t spot all AI media but finds OpenAI marks

SynthID is an invisible watermark that Google introduced in 2023. Instead of adding a label to a file’s metadata, which is easy to strip out, it hides a signal inside the content itself: in the pixels of an image or video, and in the frequency patterns of audio. People can’t see or hear it, and Google says it doesn’t affect quality.

Google says SynthID has been added to more than 180 billion images and videos and 240,000 years of audio so far. Every image Gemini makes includes it.

Until now, ordinary users had no easy way to look for the mark. You could upload a picture to Gemini and ask whether it was AI-made, but many people don’t want to trust an AI chatbot’s word about AI, and some avoid the tools altogether. A plain website with a yes or no answer solves that, and it works for content from partner companies too.

The detector simply looks for that signal. Because it isn’t guessing based on how an image looks, it shouldn’t wrongly flag a real photo as fake, which is a common problem with general AI detectors. When it finds a mark, it can also name the AI tool that made the file.

Using the free watermark checker takes less than a minute:

1. Open the SynthID Detector website and sign in. You need a Google, OpenAI or Apple account.
2. Upload an image, video or audio file. Common formats such as JPG, PNG, HEIC, MP4, MOV, MP3 and WAV all work.
3. Wait a few seconds for a result saying whether a SynthID watermark was found.

There’s a daily limit of around 10 checks, which Google reportedly set to stop people from using the results to work out how to remove the watermark. If you’d rather not use the website, SynthID checks are also built into Google Search, Chrome and the Gemini app, and Gemini can scan both the picture and sound of a video.

The headline limit is simple: the detector won’t catch AI media from tools that don’t use SynthID. Meta, for example, uses its own watermark, so a fake made with Meta’s tools would pass as clean. Google says plainly that this is not a general AI detector.

A clean result is also hedged. When one tester uploaded a real photo, the tool said it was only “unlikely” to be AI-made, because all it can confirm is that no watermark was found. It also can’t tell the difference between an image fully generated by AI and a real photo that was edited with AI, and it doesn’t show which part of a file was changed.

In hands-on tests the watermark proved hard to shake off. It survived screenshots, cropping, blurring, color and contrast changes, a painting filter and even two AI images combined into one. It failed only when an image was heavily pixelated, and on a composite that was about three quarters real photo. Google admits heavy or repeated edits can weaken the signal, so use the highest-quality copy of a file you can find.

Think of SynthID Detector as one tool in a kit, not a verdict. If it finds a watermark, you can be fairly sure the file was made or edited with a supported AI tool. If it doesn’t, keep your guard up and check where the file came from, who posted it first and whether trusted news outlets show the same thing.

Other companies offer their own checkers. OpenAI runs a detector that reads its metadata labels as well as SynthID, and Meta has a web tool for its own watermark, though a 2026 report found it sometimes missed them. Anthropic and Microsoft also mark AI media but don’t yet offer consumer checking tools.

When Google opened the detector, Apple had said it would support SynthID later in the year. That has now happened. In iOS 27.2, every image made with Image Playground carries a SynthID watermark, and so do photos changed with the Extend and Reframe tools in the Photos app.

Clean Up, the tool that erases objects, is less consistent. Testing found that small removals added no watermark, while larger edits did. There’s no way to check for the mark on the iPhone itself, so you still need the SynthID Detector website.