AI images now look so natural that the eye alone often can't tell. Still, clues left inside the file reveal a lot. Here are the practical methods and their real limits.
1. Check content credentials (C2PA)
The strongest clue is a signed content credential baked into the file. Tools like ChatGPT, Firefly, and Gemini automatically record "generated with AI." For what this is, see What are C2PA content credentials?
Want to see whether an image carries this signal? The metadata checker shows the "AI / Content Credentials" field, read locally in your browser.
2. Look at software and EXIF traces
AI tools and editors often leave a software name in the metadata. A real camera photo, by contrast, carries EXIF like the model, exposure, and capture time — which AI-generated images usually lack.
3. Clues in the image itself
- Odd finger counts, garbled text or logos, repeating backgrounds
- Unnaturally smooth skin or impossible lighting
- Shadows and reflections that don't physically add up
Newer models increasingly erase these tells, though.
The limits of metadata
Crucially, metadata can be removed. Re-encoding an image or taking a screenshot strips content credentials. So "no credential" does not mean "not AI." The reverse is a strong signal, though: if a credential is present, AI was likely involved.
There is no 100% reliable test — combine the clues above. And if it is your own content, follow each platform's AI disclosure rules.