How the answer is decided
The tool checks the strongest evidence first and stops as soon as it gets a clear answer:
- Content Credentials. If the file carries a signed record saying it was created with generative AI, that settles it — no guessing needed.
- Metadata. The IPTC source type and software tags often name the generator outright.
- Pattern model. When the file says nothing (screenshots, images saved from social apps), a detection model estimates the likelihood from the pixels.
More on each layer on the AI image detector page.
What we measured on our own test files
Before writing this page we ran four files through the exact pipeline the tool uses (October 22, 2026). Two were our own — an image generated with OpenAI's model and a re-saved copy of it — and two are public test files from the C2PA project.
| File | Content Credentials | Source type | Pattern model | Verdict |
|---|---|---|---|---|
| PNG from OpenAI's image model (our demo sample) | Found — signed by OpenAI OpCo, LLC; generator “OpenAI Media Service API” | trainedAlgorithmicMedia | Not needed | AI-generated (credentials) |
| Same image re-saved as JPEG (1280 px) | None — the 21,881-byte C2PA chunk was dropped on re-encode | — | 99% AI likelihood | Likely AI-generated (model) |
| C2PA test fixture C.jpg | Found — C2PA test signing certificate | algorithmicMedia | Not needed | Computer-generated, not generative AI |
| C2PA test fixture CA.jpg | Found — but only an “opened” action, no source type | — | Runs as fallback | Credentials can't answer it alone |
Two things stood out. One ordinary re-save erased the strongest evidence completely, which is why the pattern model has to exist. And OpenAI's signature and content hashes validated cleanly, but the official reader still reports the overall state as “Invalid” because the signing certificate isn't on the C2PA trust list — a tool that shows only that word would wrongly suggest the file was tampered with.
Check it yourself: seven visual signs
A detector is one opinion. These are the things to look at with your own eyes — none is proof on its own, but several together are a strong signal.
| Where to look | What gives it away | How reliable |
|---|---|---|
| Hands and teeth | Extra or fused fingers, too many teeth, nails that melt into skin. | Medium — newer models get these right more often |
| Text in the scene | Signs, labels and screens with letters that almost read but don't. | High when present |
| Repeating texture | Hair, grass, brick or fabric that tiles in a too-regular pattern. | Medium |
| Lighting logic | Shadows that point different ways, reflections that don't match the scene. | Medium |
| Accessories | Mismatched earrings, glasses that merge into the face, asymmetric collars. | Medium |
| Background people | Faces in the crowd that dissolve into smudges. | High when present |
| Skin and focus | Uniformly smooth skin with no pores; everything equally sharp or equally soft. | Low — filters do this too |
Tip: after a check, drag the slider on the result to view the image's noise residual. Real camera photos show fine grain everywhere; generated images often look smoother or patchier. It's a viewing aid, not part of the verdict.
Check the context, too
- Where did it first appear? A reverse image search can find an older original — or show the image only exists on one anonymous account.
- Does the story fit?A “photo” of a breaking event with no other angles, no other witnesses and no named photographer deserves suspicion.
- Who benefits? Images attached to money requests, refunds or outrage bait are the ones most worth checking.
Why you might get “inconclusive”
Heavily compressed, tiny or screenshotted images carry less signal, and a real photo edited with generative fill can look partly synthetic. When the model lands in the middle, we say so rather than pick a side. Try to get the original file — the one sent as a document rather than a compressed photo usually keeps its metadata.