Does metadata show if an image is AI?

Yes, as long as the metadata is still in the file. ChatGPT, Adobe Firefly and other services write a C2PA manifest or the IPTC value trainedAlgorithmicMedia into the image, and this tool reads both. Once the metadata is gone, the file alone cannot answer the question.

Is this image AI-generated by Google?

Google checks that itself, in the Gemini app. Images from Google's AI services carry the invisible SynthID watermark in their pixels, and only Google can read it. This tool shows Google's C2PA manifest and IPTC label when the file still has them.

A valid signature is hard to fake, but anyone can strip the manifest or replace it with an unsigned one. This tool does not verify the signature, it shows what the file contains. To check the signature, use contentcredentials.org/verify.

Does a screenshot remove AI metadata?

Yes. A screenshot is a new file without the original's metadata, so this detector finds nothing in it. An invisible watermark in the pixels can survive the screenshot, though.

Why is there no AI probability score?

Because the score failed our own test. The method this tool used before rated ten out of twelve real camera photos as AI, some higher than two genuine AI images. A number that accuses real photos is worse than no number.

Does this tool upload my image?

No. Your browser reads the file locally, with no upload and no model download. Once the page has loaded, the check works offline too.

AI Image Detector: Check C2PA and AI Labels

How does the detector recognise an AI image?

It reads the provenance record that AI services write into the file. The main one is the C2PA manifest, a digitally signed block defined by the Coalition for Content Provenance and Authenticity. PNG files keep it in a caBX chunk, JPEG in APP11 segments and WebP in a C2PA chunk. It names the program that made the image, the time it did so and the source type.

The source type comes from the IPTC digital source type vocabulary. trainedAlgorithmicMedia means a trained model produced the image. compositeWithTrainedAlgorithmicMedia means someone edited an existing image with AI, for example with generative fill. An XMP block can carry the same value without any C2PA manifest.

Local generators work differently. Stable Diffusion interfaces save the prompt, seed and sampler in a PNG text chunk called parameters, and ComfyUI stores the whole workflow as JSON. The tool recognises both and shows you the prompt.

What does a real ChatGPT image contain?

We checked an image exported from ChatGPT on 29 September 2026. The file carried a C2PA manifest of about 21 KB. From it the tool reads the generator “ChatGPT (gpt-image)”, the source type AI-generated and the time 17:17 UTC. It also finds the note that an invisible watermark was embedded. The certificate inside the manifest names OpenAI as the signer.

Why do so many images show “No provenance data”?

Because the data gets lost easily. A screenshot is a new file without it, messengers and many social networks strip metadata on upload, and some photo editors drop it when you save. In our test the ChatGPT export had its manifest, while a second version of the same picture, saved again for LinkedIn, had no metadata at all.

So “No provenance data” does not clear an image. It only means the file says nothing about how it was made.

Why did the percentage score go away?

Until September 2026 this tool also ran a neural network that looked for AI patterns in the pixels. We tested it on twelve real photos from Samsung, Apple and Fujifilm cameras. Ten of them scored above 70 percent, which the tool reported as “AI-generated”, and the highest reached 85 percent. The two genuine AI images in the same test scored 73 and 77 percent. You cannot decide anything with a number like that, so we removed it.

What about SynthID?

SynthID is Google’s invisible watermark for images from its AI services. It lives in the pixels, so it survives cropping and recompression that would wipe any metadata. Only Google can read it: the Gemini app can check an image for SynthID, and that check only recognises content from Google’s own tools. When a C2PA manifest notes a watermark, from any provider, this tool shows that note.

Do AI images have to be labelled?

In the EU, yes. Article 50 of the AI Act applies from 2 August 2026. Providers of AI systems must mark generated images in a machine-readable way, and anyone who publishes a deepfake must disclose that it is artificial. In practice the machine-readable mark is the data this detector reads: a C2PA manifest and an IPTC source type, often backed by a watermark in the pixels.

Which tools go with it?

The AI image cleaner shows the provenance data of many images at once and removes it without touching the pixels. Remove metadata also handles PDF files and lets you keep individual fields. The EXIF viewer shows camera, lens, capture time and GPS position of a photo.