Start with the original file
Metadata only survives in the file that left the generator or the camera. Screenshots, re-saved copies and images downloaded from social networks are usually re-encoded, which removes most or all of it. If you can, ask for the original file — the one exported from the app or sent “as a file” rather than as a compressed photo.
See also: Which Apps and Social Networks Strip Photo Metadata? (And When Your GPS Still Leaks)
The metadata signals that reveal AI images
From strongest to weakest, these are the signals worth checking:
| Signal | Where it lives | What it tells you |
|---|---|---|
| C2PA manifest from an AI service | APP11 (JPEG), caBX (PNG), uuid (HEIC) | Signed statement of which tool generated or edited the image |
| IPTC DigitalSourceType | XMP metadata | Declared origin, e.g. trainedAlgorithmicMedia |
| Prompts and parameters | PNG text chunks, EXIF UserComment | Generated by a local tool such as Stable Diffusion or ComfyUI |
| Tool names | XMP CreatorTool, EXIF Software | Which app created or last saved the file |
Signal 1: C2PA Content Credentials
Content Credentials are a digitally signed record attached by tools that support the C2PA standard. The claim generator names the app or service that produced the image, the c2pa.actions list records steps such as “created” or “edited”, and a digitalSourceType value can declare AI generation outright.
Because the record is signed, verifiers can also tell whether the image was modified after signing. A manifest from an AI service is the strongest metadata evidence you can get.
See also: What are C2PA / Content Credentials?
Signal 2: IPTC DigitalSourceType
Many generators and editors write a standard IPTC field that describes how the content was made. The values you will see most often:
- trainedAlgorithmicMedia — created by a generative AI model.
- compositeWithTrainedAlgorithmicMedia — a real photo with parts generated or replaced by AI (for example Generative Fill).
- algorithmicMedia — created by an algorithm without training data, such as procedural graphics.
- digitalCapture — captured by a camera, with no generative content declared.
Signal 3: Prompts, parameters and tool names
Local generators such as Stable Diffusion and ComfyUI store the prompt, seed, model and workflow as text inside PNG files. Other tools leave their name in XMP CreatorTool or EXIF Software. Any of these is a clear sign of AI involvement — and a clue to exactly how the image was made.
See also: What Metadata Do AI Image Generators Add? DALL·E, Firefly, Midjourney, Gemini and Stable Diffusion · Remove Stable Diffusion prompts and ComfyUI workflows from PNG files
How to check an image in seconds
- Drop the file into the tool at the top of this page. The check runs entirely in your browser; the image is never uploaded.
- Read the verdict in the “AI & provenance” panel: “AI signals detected”, “Needs a closer look” (Content Credentials without an explicit AI declaration) or “No AI signals”.
- Review the listed signals and their sources, then open “All metadata & file details” for the raw values.
What to do when there is no metadata
Missing metadata is not evidence either way — editing apps, messaging apps and social networks strip it all the time. Other checks you can use:
- Reverse image search to find the earliest copy and its original context.
- The provider’s own watermark checker — for example Google’s tools for SynthID — where one exists.
- Visual inspection: hands, text, reflections, repeating textures and inconsistent lighting are common weak points, though modern models make fewer mistakes.
- The context: who posted it first, whether reliable sources confirm the scene, and whether the account usually posts AI content.
See also: SynthID and Invisible AI Watermarks Explained: What Metadata Cleaning Can and Can’t Do
How reliable are AI image detectors?
Pixel-based AI detectors estimate a probability; they do not prove anything. Results vary between detectors, drop after compression or resizing, and can flag real photos as AI. Treat them as one input among several — and give more weight to signed provenance and to the original context.
Questions about checking images for AI
Does “no AI signals” mean the photo is real?
No. It only means the file carries no AI metadata. Metadata is easy to lose or remove, so treat a clean result as “unknown”, not as proof of authenticity.
Can I check images downloaded from Instagram or Facebook?
You can, but platforms usually strip metadata from the published copy, so the result is often empty. The platform’s own “AI info” label, if present, is a separate signal.
Can metadata be faked?
Plain EXIF and XMP fields can be edited with common tools. C2PA manifests are signed, so tampering is detectable — which makes them the more trustworthy signal.
How can I tell if an image was made with ChatGPT?
Images from OpenAI’s image generation carry C2PA Content Credentials that name OpenAI. If the original file still has them, the report shows them under “AI & provenance”.
Can I check Midjourney or Stable Diffusion images?
Yes. Midjourney downloads may contain the prompt and an AI source type, and Stable Diffusion PNGs usually contain a “parameters” chunk — both are flagged in the report.
Is the checker free and private?
Yes. It is free, needs no account and analyzes the file locally in your browser without uploading it.