Google Turned “Check Whether This Image Is AI-Generated” Into a Website

On October 7, Google brought the SynthID Detector to the public web. The address is short: synthid.com. After you log in, you can drag in an image, a video, or an audio clip, and a few seconds later it will tell you: watermark detected, or not detected.

In the past, you had to ask Gemini for help with this. You uploaded a file, then watched the model describe in natural language what it saw; getting a crisp “yes” or “no” was a struggle. Now Google has extracted the judgment into a standalone entry point that only requires logging in, with no need to understand the model's wording.

180 Billion Images, and 240,000 Years of Audio

SynthID is not new. Google put it to use in its own generative models back in 2023. The method is to encode a signal invisible to the human eye and inaudible to the human ear directly into pixels and audio waveforms, rather than stuffing it into metadata that can be cleared with one click. This matters: metadata is like a sticky note on the back of a picture frame—after copying and pasting, screenshotting, or re-saving, it is often gone, while a signal written into the pixels is carried along with the image. According to figures Google has published, since launching in 2023 it has marked more than 180 billion images and videos, plus audio content equivalent to 240,000 years. The audio figure sounds exaggerated, but it actually includes everything produced by Lyria music generation and the NotebookLM podcast feature.

The watermark itself has also been made resistant to modification. According to Google DeepMind's technical description, image and video watermarks are written in at the moment of generation, and after cropping, filters, frame-rate changes, and lossy compression, they still have a relatively high chance of being detected; audio watermarks are hardened against added noise, MP3 compression, and speed changes. Text generated by Gemini takes another route: it embeds a signal by fine-tuning the model's scoring distribution for each candidate token, with no visible change to the words you see.

The key change in this opening is the partners. The previous detector only recognized Google's own watermarks; an image generated by ChatGPT, even if it also carried SynthID, would still not be flagged in Google's tool. Now this limitation is gone: the detector supports OpenAI, NVIDIA, and South Korea's Kakao at the same time, and Apple has promised to join later.

All told, one upload can check for generation traces from several companies at once, so you no longer have to go to each company's own verification page one by one. Google has also connected this capability into Search, the Gemini app, and Chrome; together, these three handle more than 1 million verification requests per day.

Why You Can’t Run Unlimited Checks

The tool itself is free, but there are limits. Each account gets only about 10 checks per day for images, video, and audio, and you must sign in with a Google, OpenAI, or Apple account to use it. Google engineers have explained the reason: access is restricted to prevent people from using detection results to reverse-train a “watermark-removal” tool. The quota number is deliberately vague; if you upload a batch of very similar images in a row, the system may lock you out earlier, because it looks like you're probing for some kind of bypass.

One more detail is worth noting. Uploaded files are processed in real time and deleted immediately after the result is produced, but the file's digital signature is retained for 24 hours to track the day's quota.

“Not Detected” Does Not Mean “Not AI-Generated”

This is the most important thing to be clear about with the entire tool, and Google says it in its FAQ: it is not a general-purpose AI detector. It can only identify content produced by companies that use SynthID technology.

In reality, a great deal of content naturally falls outside what it can judge. Microsoft and Meta use their own watermarking standards, and neither is on Google's partner list; this year Meta also had an incident where its own detector missed AI images that its own system had cropped. More important are open-weight models, which anyone can run on their own machine; the content they generate does not even have the most basic metadata tags.

So the correct reading of “no watermark detected” is “unknown origin,” not “this is real.” The file may come from a model that hasn't adopted SynthID, or it may simply have been cropped, compressed, or re-recorded so heavily that the signal was worn away.

Google's other approach is just the opposite: rather than labeling AI content, label real content. Cryptographic methods such as C2PA sign provenance information in at the capture device, and Pixel phones are one of the few products already doing this. One labels fake content, the other labels real content; the two paths each cover one end.

For ordinary users, the site's practical value is that it is a unified entry point. When you see a suspicious image, drop it in for a check first, then make a judgment using the source, time, and context—more reliable than looking at the result alone. It is worth noting that it supports a fairly wide range of formats: images include JPG, PNG, WEBP, and HEIC; video covers MP4, MOV, and WEBM; audio includes common types such as WAV, MP3, and FLAC. The vast majority of files you encounter day to day can be uploaded directly.

In Chinese-language contexts, debates over “was this image AI-drawn?” have not stopped for the past six months, and finished images from stock-photo sites are increasingly hard to distinguish from illustrations that appear on your phone. Having one more place to check is always a good thing, as long as you don't treat it as a verdict. The real habit to build is this: watermark detection gives you a single-point signal. When it says “yes,” it is fairly credible; when it says “no,” it has said almost nothing. Whether an image is trustworthy ultimately depends on who took or generated it, where it was published, and whether a second party can corroborate it. A tool can help you rule some things out, but it cannot draw the conclusion for you.