The short answer
A metadata check proves an image came from a generator. It never proves the opposite. Absence of a label means the label went missing somewhere, and that happens constantly.
- Label present: strong evidence An IPTC Digital Source Type value, an XMP generator field, or a signed C2PA manifest names the origin directly. Files rarely gain those fields by accident.
- Label absent: no evidence either way Screenshots, social uploads, messaging apps, and most editors drop provenance fields. A clean file proves only that the file is clean.
What a real AI label looks like
Three separate systems write these fields, and a generated file often carries more than one.
| Signal | Where it lives | What it tells you |
|---|---|---|
| IPTC Digital Source Type | IPTC or XMP block | A controlled value such as trained algorithmic media, composite synthetic media, or digital capture. IPTC publishes the vocabulary, so the values mean the same thing everywhere. |
| XMP generator fields | XMP packet | The tool name, the model, sometimes the prompt text and a job identifier. |
| C2PA Content Credentials | A signed manifest inside the file | Origin plus edit history, cryptographically signed. The C2PA specification defines the format. |
| Software and comment tags | EXIF | Weaker, but a Software value naming a generator still counts as a lead. |
Check a file in three steps
- Get the original file Ask for the file as sent, not a screenshot and not a social download. A re-saved copy has already lost the fields you want to read.
- Open it in the viewer Load it in the metadata tool. Nothing uploads. The AI and origin panel shows the declared generator and every provenance field the file holds.
- Read the result literally Fields present means the file says so. Nothing found means the file says nothing. The tool does not convert silence into a verdict, and neither should you.
Why pixel detectors disagree with each other
An AI image detector looks at the picture and scores how synthetic it appears. That score comes from a model trained on examples, so it inherits every gap in that training set. Run one photograph through three detectors and you can collect three different answers.
The failure modes are predictable. Heavy compression, a strong denoiser, a beauty filter, or an upscaled crop all push a real photograph toward a synthetic score. A generated image passed through a screenshot and a social upload often lands on the safe side. Neither result carries any proof.
Watermark systems that live in the pixels sit in a third category. Google DeepMind describes SynthID as an imperceptible signal embedded in the pixels themselves and detected by SynthID technology. Only a matching detector reads it. Vaultaire never inspects pixels, so it reports a SynthID string only when the metadata mentions one, and otherwise says the pixel check did not run.
How the labels disappear
Most label loss has nothing to do with anyone hiding anything.
- Screenshots A screenshot is a new picture of a screen. It carries your device's metadata and none of the original file's.
- Social platforms and messaging apps Uploads get re-encoded to save bandwidth. The re-encode keeps the pixels and throws away most metadata blocks.
- Ordinary edits Crop it, resize it, or export it as a different format and many editors write a fresh file with only the fields they care about.
- Deliberate removal A metadata remover clears the fields on purpose. That is the same one-way problem from the other side.
AI image checking FAQ
Can you tell if an image is AI generated?
Sometimes, and only in one direction. When the file carries an AI label such as the IPTC Digital Source Type value trained algorithmic media, a generator name, or a signed C2PA manifest, you have a definite answer. When the file carries none of those, you have no answer at all, because saving and sharing strip those fields routinely.
Are AI image detectors accurate?
Pixel detectors return a probability, not a fact, and they misfire in both directions on ordinary photographs. Use them as a hint and never as evidence.
Does a screenshot keep AI metadata?
No. A screenshot creates a new file from what the screen showed. It carries the metadata your device writes and none of the original file's AI labels or C2PA record.
Does Instagram remove EXIF data?
Every large social platform re-encodes uploads and drops most of the original metadata, AI provenance fields included. A picture saved from a social feed almost never carries what the generator wrote.
Which generators can the report name?
When the file records it, the report names Adobe Firefly, AUTOMATIC1111 or Stable Diffusion WebUI, ChatGPT or OpenAI, ComfyUI, DALL-E, Flux, Google Imagen, InvokeAI, Midjourney, and Stable Diffusion. It reports the text the file holds and never guesses beyond it.
Does the file leave my device?
No. Parsing runs in a Web Worker in your browser. No upload, no account, no server copy.