A photo that looks real proves nothing in 2026. Modern image generators handle hands, faces, and short text well enough that visual inspection alone often fails, and that is why watermark detection has become a useful check. Google's SynthID Detector, opened to the public in October 2026, lets anyone upload an image and ask whether it carries Google's invisible SynthID watermark.

SynthID is a watermark embedded directly into the pixels of content made or edited with participating Google AI tools. Unlike a visible label, it survives ordinary edits such as cropping and compression in many cases, though it is not indestructible. The detector scans the file for that signal and reports where it appears, if it appears at all.

Using the detector is straightforward. First, make sure you are on Google's official SynthID Detector page by searching for it directly, and avoid lookalike sites asking you to upload files. Upload the original file whenever possible rather than a screenshot or a forwarded copy, because re-saving and heavy compression can weaken or strip the watermark. Supported image formats include JPG, PNG, WebP, and several others.

Once the scan finishes, read the result carefully. A detected watermark means the file, or part of it, was produced or edited with a participating Google tool. It does not necessarily mean the entire image is AI-generated: a real photo edited with an AI tool can carry the watermark in the edited region only. The detector may show which areas carry the signal, so treat the result as a map, not a single yes-or-no verdict.

A result showing no watermark is much weaker evidence. SynthID only covers Google-origin content, so images from other generators, or images whose watermark was stripped by screenshots, re-encoding, or aggressive compression, will come back clean. A clean result means 'no Google watermark found,' not 'this photo is authentic.'

That limitation is why the detector works best as one check among several. Before drawing conclusions, look for the image's source: who posted it first, whether a reputable outlet or photographer is attached to it, and whether a reverse image search finds earlier copies. Check for platform labels such as 'AI info' where they exist, and inspect the file's Content Credentials if they survived, since these record how a file was created and edited.

There are also practical limits to keep in mind. Uploading someone else's private photo to any detector raises privacy questions, so only check images you have a legitimate reason to verify, such as something you are about to share or publish. And never use a detector result to accuse a real photographer of faking work: detection tools produce false results, and a single score is not proof.

A sensible workflow looks like this. First, ask where the image came from and whether the source is accountable. Second, check for Content Credentials and platform labels. Third, run a reverse image search for earlier versions. Fourth, use the SynthID Detector on the best-quality original you can find. If all of those checks come back empty, label the image unverified rather than real or fake, and avoid sharing it as evidence of anything.

The public detector is not limited to still images. It also accepts video and audio files, checking for SynthID signals in segments of the recording. This matters because manipulated clips now spread faster than photos, and a partial watermark can show which part of a clip was generated or edited even when the rest is genuine. The same caution applies: a clean result does not clear a clip, and a positive result only covers the watermarked portion. For video, also watch for lip-sync drift, unnatural blinking, and voices without breaths, which are weaker but still useful supporting signals alongside the watermark check.

Watermark detection is a genuine step forward, but it only covers part of the picture. The strongest habit is still the oldest one: verify the claim around an image before trusting the image itself.