How to Check Whether an Image Has a SynthID Watermark

You cannot see a SynthID watermark by looking at an image, and you cannot confirm one by squinting at the corners. It is an imperceptible signal embedded directly into the pixels, so checking for it means running the file through a system that can read that signal. In 2026 the reliable way to do that is Google’s own tooling. This guide walks through the methods that actually work, then explains what a result does and does not prove, which matters more than most people expect.

For background on the signal itself, see what SynthID is. If your goal is editing rather than verifying, the AI watermark remover overview covers that side. This page is about detection.

#The methods that actually check for SynthID

Because the watermark can only be verified with Google’s private keys, the trustworthy options all route through Google. Third-party “SynthID detectors” cannot cryptographically confirm the mark; at best they estimate a likelihood. Here are the practical methods, from most authoritative down.

#Using the SynthID Detector and Gemini

Google’s SynthID Detector portal is the closest thing to an authoritative check. You upload the file and it reports where the watermark is likely present, often highlighting the specific regions that carry it. The Gemini app offers a more casual route: upload the image and ask whether it was generated or edited by Google AI. Both lean on the same underlying detection, so treat them as two front doors to one answer rather than two independent verdicts.

Neither is a public API. There is no endpoint you can call at scale, and there is no offline tool that reads the watermark, because doing so requires keys only Google holds. Any service claiming to detect SynthID without Google is inferring, not verifying.

#What a result does and does not prove

This is the part that trips people up. SynthID only marks content from Google AI and its partners, and since 2026 that circle includes OpenAI’s ChatGPT and DALL·E images. It does not mark output from every model in existence.

• A positive result is strong evidence the image came from a participating AI system.
• A negative result does not prove the image is human-made. It could be from a model that does not use SynthID at all.
• Highlighted regions indicate where the signal is strongest, not the only places it exists.
• Heavy edits can suppress the signal, so a faint or absent read on a modified file is not conclusive either way.

#How to get the most accurate read

The watermark is designed to survive cropping and light compression, but you still get the cleanest signal from the least-touched file. A few habits improve your odds.

1. Use the original, unedited file whenever you have it, since re-exports and filters weaken the signal.
2. If you can only screenshot, crop tightly to the image and avoid rescaling.
3. Avoid stacking heavy edits before checking; each one reduces what the detector can read.
4. Treat provenance separately: check contentcredentials.org/verify for C2PA metadata, which is independent of the pixel watermark.

If cropping is on your mind because you are wondering whether it strips the mark, the short answer is that it usually does not. See whether cropping removes SynthID for the details. And if you actually own the image and need to repair a visible mark, the removal walkthrough covers that workflow.