Can Gemini or ChatGPT tell if an image is AI?

It is a fair thing to ask the chatbot you already use: can Gemini tell if an image is AI? The short answer is that Gemini can check an image for SynthID, a hidden watermark Google adds to media made with its own tools, and it will tell you when it finds one. The catch is what happens when it does not.

This article explains what SynthID is, how to run the check in Gemini, whether ChatGPT can do the same, what watermark checks miss, and how to handle an image from any other generator. The theme throughout: a watermark check, a probability detector and proof at capture answer three different questions, and it helps not to confuse them.

What SynthID is and how the check works

SynthID is a watermark Google embeds into images, audio, video and text produced by its AI tools. It is woven into the content itself rather than sitting in a metadata field, so it can survive ordinary editing like cropping or recompression. When you ask Gemini whether an image is AI, it looks for that watermark and then writes a short explanation of what it found.

The scope is the thing to understand. According to Google, SynthID identifies content made with Google's own AI tools. The Register reported when Gemini gained this image check that it cannot tell whether a rival generator produced a picture. So a SynthID result confirms Google origin when it is present, and stays silent about everything else.

How to check an image with Gemini, step by step

Open the Gemini app or website, start a new chat, and upload the image you want to check. Ask plainly whether it was made with Google AI, or whether it contains a SynthID watermark. Gemini inspects the file and replies with whether all or part of it appears to have been generated with Google's tools.

Read the answer carefully, because the wording matters. A positive result is meaningful: the watermark is there, so Google AI was involved. A result that finds no watermark is not a clean bill of health. It means only that Gemini did not detect a Google watermark, which is exactly what you would see for a real photo, for a screenshot, or for an image made by a tool that is not Google's.

Can ChatGPT check images too?

ChatGPT is not a general image detector, but OpenAI has built provenance into its own images. Per OpenAI's help documentation, images generated in ChatGPT, Codex and the OpenAI API carry both C2PA content credentials and a SynthID watermark. PetaPixel notes the watermark is embedded in the pixels, so it can survive even when the metadata is stripped. OpenAI also runs a public checker at openai.com/verify that looks for a SynthID watermark or a trusted OpenAI C2PA manifest and reports whether an image carries OpenAI provenance.

The same limit applies as with Gemini. These tools confirm their own origin when the signal is present, but a negative result does not prove an image is real. OpenAI itself says a missing tag does not mean an image did not come from ChatGPT, because metadata can be removed. Asking ChatGPT to eyeball a random photo and judge whether it is AI is not a reliable test.

What watermark checks miss

Watermark checks are powerful but narrow. They miss any image from a generator that does not watermark its output, and plenty do not. They can miss an image whose watermark was stripped or degraded by heavy editing, screenshotting or re-encoding. And they say nothing about a genuine photo of a real scene, which carries no watermark because no AI touched it.

Metadata is even easier to lose. The C2PA tag on an AI image lives in the file's metadata, and you can inspect it yourself with a free metadata viewer, but social platforms and messaging apps routinely strip metadata on upload. So a shared image arriving with no watermark and no metadata tells you almost nothing on its own.

How to check an image from any generator

When the watermark check comes back empty, fall back to a tool that does not depend on cooperation from the generator. A free AI image detector looks at patterns in the pixels and returns a probability that the image was AI generated, whatever tool made it. It is not tied to Google or OpenAI, so it covers the images their checks cannot.

Use the two together. A watermark check is a quick confirmation of Google or OpenAI origin, and a probability detector is your fallback for everything else. Neither is a verdict, so treat a high detector score as a strong reason for caution and a low one as reassurance, not proof.

Watermark, detector, proof at capture: three different answers

These three tools answer three different questions. A watermark says a specific company's AI made this. A detector estimates how synthetic a finished file looks. Neither can prove that a real photo is real, which is the question that matters when you need your own images to hold up.

That is where proof at capture comes in. Instead of judging a file after the fact, it records authenticity at the moment of capture: binding the photo to a SHA-256 fingerprint, a server timestamp and a GPS position from the phone as it is taken, so anyone can verify later that it has not changed. Watermarks and detectors look backward at an existing file. Proof at capture starts the record from the first moment, which is the only way to show a photo is genuine rather than guess at it.

FAQ

No watermark found? Check anyway.

Danaya's free detector reads the pixels, not a watermark, so it works on images from any generator. Upload one and get a probability in seconds.

Check an image free

Keep reading