Ten checks, strongest first
- 1
Look for Content Credentials
ChatGPT, Gemini, Adobe Firefly and Microsoft Designer sign their images with a C2PA manifest. If it is there, it says how the image was made.
- 2
Read the IPTC source type
The value “trained algorithmic media” in the image's XMP metadata is the industry label for AI-generated.
- 3
Find generator settings in the file
Stable Diffusion and ComfyUI images often still carry the prompt, seed and model name in the PNG.
- 4
Run SynthID
Google's SynthID Detector reads the invisible watermark in images from Google, OpenAI and others, and survives screenshots.
- 5
Reverse image search
Find where the image first appeared. An account that posts AI art, or an earlier version with a watermark, answers the question.
- 6
Check hands, teeth and ears
Extra or fused fingers, too many teeth and mismatched earrings are still the most common tells, though the newest models make fewer.
- 7
Read every bit of text
Signs, labels and logos that are almost-words, or letters that melt into each other, point to a generator.
- 8
Follow the light
Shadows that fall in different directions, reflections that do not match the scene, and glasses or eyes that reflect nothing coherent.
- 9
Look at the background
Repeated faces in a crowd, railings and windows that do not line up, and objects that merge into each other.
- 10
Judge the texture
Waxy skin without pores, hair that dissolves at the edges and an even, rendered sharpness across the whole frame.
Run every check at once
The first four checks are proof straight from the maker; the rest are the tells that give generated images away. Together they catch far more than any one of them.
The Gireta detector above runs them in one go, in seconds, and shows you exactly which ones it found.
What each AI image generator leaves in its files
Most of the big generators now label their images. The label survives as long as nobody screenshots, crops or re-saves the file.
Screenshots, social apps and most image editors strip all of this. A missing label does not mean a real photo.
More ways to check
Images
Video
Make AI images that stand out
Gireta puts Nano Banana, GPT Image, Seedream and every top image model in one studio, with one balance.
Questions
What is the fastest way to tell if a photo is AI?
Drop it into an AI image detector that reads Content Credentials, such as the Gireta detector on this page. If the file carries a manifest or IPTC label, you get a definite answer in a second. If not, look at hands, text and reflections.
Can you tell if an image is AI from its metadata?
Often. ChatGPT, Gemini and Firefly write Content Credentials, Midjourney committed to the IPTC source type, and Stable Diffusion and ComfyUI store their settings in the PNG. Screenshots and social apps strip all of it.
Do AI images still have bad hands?
Less often than in 2023. The newest models draw hands well most of the time, so a good hand proves nothing. A bad one is still a strong clue.
Do you keep my image?
No. The file is read in your browser and never uploaded. For the visual check, a small preview (at most 1024 pixels wide) is sent, inspected and discarded.
Sources
- IPTC: digital source type for generative AI
- OpenAI: C2PA in ChatGPT images
- Google: SynthID Detector
- ComfyUI: workflow metadata in images
Last updated October 19, 2026.