Check what the file says about itself first: many AI tools add Content Credentials or an AI label in the metadata. Next, test for an invisible watermark such as Google's SynthID. Then reverse-search the image to find where it first appeared. Visual clues like odd hands or garbled text come last, because they're the least reliable.
1. Read the file's own labels
Many AI tools write a note into the file that says made with AI. There are three places to look:
- Content Credentials (C2PA): a signed record from the company that made the image. Used by ChatGPT, the Gemini app, Adobe Firefly, Microsoft's Image Creator and others. You can verify the signature at verify.contentauthenticity.org.
- The AI label in metadata: a standard IPTC field called digital source type, set to trainedAlgorithmicMedia for AI images.
- Saved prompts: Stable Diffusion tools often store the full prompt and settings.
Our AI image checker reads all three in your browser. If it finds a label, you have your answer. If it finds nothing, move on: labels are lost whenever an image is screenshotted, re-saved or uploaded to a site that strips metadata.
2. Test for an invisible watermark
Some generators hide a watermark in the pixels themselves. Google's SynthID is the biggest: it's used for Google's own images and has been adopted by other companies, including OpenAI. It survives cropping, filters and screenshots far better than metadata does.
Only Google's detector can read it. Upload the image at synthid.com, or ask the Gemini app whether an image was made with Google AI. Read SynthID explained for what a result means.
3. Find where it came from
Run a reverse image search with Google Lens, Bing Visual Search or TinEye. You're looking for:
- The earliest copy. An AI artist's page, a stock site's AI section or a known AI account answers the question.
- Fact-checks. Viral images are often already checked by news agencies.
- The original photo, if someone edited a real image.
4. Look closely, but don't rely on it
Visual clues still help sometimes: hands and fingers, text on signs, jewellery and ears, reflections that don't match, background objects that melt into each other, and lighting that comes from two directions. But current generators get these right more often than not. Treat any visual clue as a reason to dig further, not as proof.
5. Use a detector as a last opinion
Classifier detectors judge how an image looks and give a probability. They're useful when nothing else turns up, but read the score as an opinion. They can flag heavily edited real photos and miss AI images that were compressed or resized.
Checking a video
The same steps work for video. Our AI video checker reads Content Credentials, AI labels and location from MP4 and MOV files, and SynthID Detector checks video frames.
Putting it together
| What you found | What it means |
|---|---|
| Content Credentials or an AI label saying AI | Almost certainly AI-generated, by the named tool |
| SynthID detected | Made or edited with a tool that uses SynthID |
| An earlier copy credited to an AI artist | Very likely AI |
| No labels, no watermark, real photo source found | Very likely a real photo |
| Nothing found either way | Unknown. Absence of a label isn't proof of anything |