The file says it is AI
A source type of trainedAlgorithmicMedia or a saved prompt means a tool marked the file as generated. Platforms that read metadata will likely label it.
Reads Content Credentials (who signed it, which tool made it, whether AI was declared), IPTC AI tags and saved prompts. Not a pixel detector.
Inspect supported file records without editing the original.
It reads the records inside an image file, not the pixels.
An AI metadata checker opens the hidden records stored with an image and looks for the tags AI tools write. That means a C2PA Content Credential, an IPTC DigitalSourceType of trainedAlgorithmicMedia, or prompt text saved by Stable Diffusion and ComfyUI.
Many generators label their output this way. ChatGPT and Adobe Firefly attach Content Credentials. Meta AI adds IPTC metadata. Stable Diffusion tools save the prompt in the PNG.
Instagram, Facebook and TikTok read the same records to decide when to show an AI label. Checking first tells you what a platform will find.
| ChatGPT / DALL-E / Sora | C2PA signed by OpenAI, declaring AI generation |
|---|---|
| Adobe Firefly | C2PA declaring AI generation |
| Google Gemini / Imagen | SynthID in the pixels; some exports also carry C2PA |
| Meta AI and others | IPTC DigitalSourceType = trainedAlgorithmicMedia |
| Stable Diffusion / ComfyUI | Prompt and settings in PNG text |
Upload, read the report, then decide what to do with the file.
Drop up to 30 JPG, PNG, WebP, AVIF or HEIC files. They are read in your browser and never uploaded.
AI findings come first: generation fields, prompt text and any Content Credential. Every other metadata group follows, with decoded fields.
Keep the file as it is, validate the credential in the C2PA checker, or make a clean copy with the AI metadata remover.
Four kinds of record can mark an image as AI-made. They are not equally strong.
| Signal | Where it lives and who writes it | How much it tells you |
|---|---|---|
| C2PA Content Credentials | JUMBF block in JPG, caBX chunk in PNG, C2PA chunk in WebP. Written by ChatGPT, Adobe Firefly, Photoshop and some cameras. | Signed and tamper-evident. Found here, validated in the C2PA checker. |
| IPTC DigitalSourceType | XMP field set to trainedAlgorithmicMedia (fully AI) or compositeWithTrainedAlgorithmicMedia (AI edits to a real photo). Used by Meta AI, Adobe and others. | A clear label, but unsigned. Any editor can add or remove it. |
| PNG prompt text | PNG text chunks named parameters, prompt, workflow or negative_prompt. Written by Automatic1111, Forge and ComfyUI. | Strong hint of local generation. Plain text that is easy to delete. |
| Generator fields in XMP | Prompt, Workflow or GenerationParameters fields some apps write into XMP. | Useful context. Unsigned and easy to change. |
The checker does not look at pixels. It cannot see SynthID or other invisible watermarks, and it does not give an AI likelihood score.
JPG, PNG, WebP, AVIF and HEIC still images. Up to 30 files, 15 MB each. Images up to 50 megapixels; animated PNG and WebP are not supported. Free, no account.
A finding is a fact about the file. It is not a verdict on how the image was made.
A source type of trainedAlgorithmicMedia or a saved prompt means a tool marked the file as generated. Platforms that read metadata will likely label it.
A C2PA container is present. Open the C2PA checker to see who signed it and whether the signature is still valid.
Screenshots, re-saves and social media downloads strip metadata. An empty result does not prove a photo is real.
Generative Fill and other AI edits can mark a real photo as a composite. One AI feature is enough to add the tag.
Answers about results, privacy and limits.
Screenshots, re-saving and most social media downloads strip it. No metadata does not mean the image is human-made.
Unsigned fields can be edited by anyone. Content Credentials are signed, so changes show up as an invalid signature.
Not a pixel one. It reads what the file says about itself. For invisible watermarks such as SynthID, see the SynthID flow.
Any that write Content Credentials or IPTC AI tags, such as ChatGPT, Firefly and Gemini, plus Stable Diffusion and ComfyUI prompts in PNGs.
No. Metadata shows what a tool wrote into the file. A signed Content Credential is the strongest evidence, but unsigned fields can be added, edited or removed by anyone.
ChatGPT and DALL-E, Adobe Firefly and Photoshop attach C2PA Content Credentials. Meta AI adds IPTC metadata. Stable Diffusion WebUI, Forge and ComfyUI save the prompt in PNG text chunks. Midjourney output varies, so check your own file.
An AI feature touched it. Generative Fill, AI erase, background removal and upscaling can write a composite source type or a Content Credential into an otherwise real photo.
No. SynthID is hidden in the pixels, not the metadata, and removing metadata does not remove it. The SynthID remover on the homepage handles that.
Yes. Meta says it reads C2PA and IPTC signals to apply AI Info. TikTok reads Content Credentials for its automatic AI-generated label.
Use the AI metadata remover to make a clean copy, or the AI-only remover to strip just the generation fields and keep camera details. Check the copy again here before you post it.
Yes. It is free, with no account, credits or trial. The file and batch limits are shown on the upload box.
No. Files are read in your browser. Save your results before you clear the batch or leave the page.
Free, in your browser, up to 30 images at once.