ExifZero How to tell if an image is AI-generated
By the ExifZero team · Updated
Short answer: start with the file itself. Many AI tools label their output in the metadata — a C2PA Content Credentials manifest, the IPTC “trainedAlgorithmicMedia” source type, or the prompt and settings — and a metadata checker reads these in seconds. If the metadata is gone, which is normal for screenshots and images saved from social networks, combine a reverse image search, a vendor watermark check and a close look at the details. No single method proves an image is real.
1. Read the image's metadata
This is the fastest check and the only one that can name the tool. Hosted generators such as ChatGPT, Adobe Firefly and Google's image models attach labels, and local tools such as Stable Diffusion and ComfyUI save the full prompt. Drop the original file into the AI image checker and look for:
- Content Credentials (C2PA) with an action whose digital source type is
trainedAlgorithmicMedia— the signer declared the image AI-generated. - IPTC Digital Source Type in XMP — the same vocabulary, written by Midjourney, Meta AI and many stock sites.
compositeWithTrainedAlgorithmicMediameans a real photo was edited with generative AI. - Generation parameters — a prompt, negative prompt, seed, sampler or model name in PNG text chunks or the EXIF comment.
- Software names — an AI tool named in the EXIF
Softwarefield or the XMP creator tool.
Use the original file whenever you can. Each re-save, screenshot or upload is a chance for the metadata to be dropped. See what each AI generator writes for a tool-by-tool list.
2. Look for Content Credentials where the image is published
Some platforms read C2PA on upload and show it to viewers, typically as a small “cr” icon or an “AI info” label. If you have the file, the C2PA viewer shows the generator, the edit actions, the ingredients and the organisation that signed the manifest. Our viewer displays the signature without verifying it; for a trust decision, also check the file with the Content Authenticity Initiative's verify tool.
3. Check for an invisible watermark
Some vendors hide a watermark in the pixels, which survives screenshots and re-compression far better than metadata. Google marks images from its models with SynthID and lets you check an image with its own tools, such as asking Gemini whether an image was made with Google AI. These detectors only recognise that vendor's watermark: a negative result says nothing about images from other generators.
4. Run a reverse image search
Google Lens, Bing Visual Search and TinEye find other copies of the image. The earliest copy often leads to the creator, a caption that says how it was made, or a fact-check. A real news photo usually appears first on a news agency or photographer's site; an AI image often first appears on an AI art community or a social account that posts generated content.
5. Look closely at the details
Current generators are good, but they still make characteristic mistakes. Zoom in and check:
- Text — signs, labels and book spines with letters that drift into nonsense.
- Hands, teeth, ears and jewellery — merged fingers, mismatched earrings, too many teeth.
- Physics — shadows that point different ways, reflections that do not match, objects that melt into each other.
- Backgrounds — faces in a crowd that dissolve, repeated patterns, architecture that does not join up.
- Texture — skin and surfaces that are too smooth or uniformly sharp, with an airbrushed look.
Treat these as hints, not proof: newer models make fewer of these errors, and heavily edited real photos can show some of them.
6. Use AI-image detectors with caution
Classifier tools estimate the probability that an image is synthetic from its pixels. They can be useful as one signal, but they produce both false positives and false negatives, struggle with compressed or resized images, and lag behind new models. Do not accuse anyone on the strength of a detector score alone.
What each method can prove
| Method | Strong evidence of AI when… | Limitation |
|---|---|---|
| Metadata check | A label, C2PA action or prompt is present | Easily lost or removed; plain fields can be edited |
| Content Credentials | A verified manifest declares AI generation | Only tools that write C2PA; stripped by many apps |
| Vendor watermark check | The vendor's detector finds its watermark | Recognises only that vendor's own models |
| Reverse image search | The earliest source is an AI account or says so | New or private images have no history |
| Visual inspection | Clear impossibilities in text, hands or physics | Subjective; modern models make fewer errors |
| Detector tools | Several detectors agree with high confidence | False positives and negatives are common |
Why “no AI tags found” does not mean “real”
Most social networks and messaging apps strip metadata on upload, screenshots never carry it, and some generators write none at all. An empty result only tells you the file makes no claim about itself. The reverse is also true: plain EXIF and XMP fields can be typed in by anyone, so an AI label in unsigned metadata is strong evidence, not proof.