Guides · 7 min de lecture
How to tell if an image is AI-generated (and when you can't)
Vajeeran Logeswaran · Updated · Sources and corrections
An image can be camera-captured, fully generated, partly edited or impossible to classify from the available file. This workflow separates provenance, detector estimates and visual observations.
1. Preserve the original file
Ask for the original download or camera file, if available. A screenshot, messaging app or social platform can remove metadata and change the pixels. Keep the received file unchanged and perform checks on a copy. Record where and when it was obtained.
An original file gives you more evidence to inspect, but the person supplying it may still be mistaken about its origin. Do not equate “original download” with “camera original.”
2. Read Content Credentials carefully
C2PA credentials can describe creation and edits in a signed manifest. Check the signature, its association with the file, the signer trust status and the actual assertions. A credential being present is different from successfully validating it.
Credentials do not prove that a depicted event occurred. A photo of a staged scene can have valid credentials. Missing credentials also do not establish AI use: credentials may never have been added or may have been removed.
3. Separate metadata from stronger provenance
Software names and generation settings in EXIF, XMP or PNG text can be useful leads. These fields can be edited or copied, so they are declarations to corroborate rather than proof. Camera make and exposure information can also be forged.
Record the exact field and value, not just “AI metadata found.” Check whether the marker describes creation, a later edit or an ingredient within a composite image.
4. Use detectors within their coverage
A pixel detector estimates patterns learned from training data. Compression, cropping and new generators can change its performance. A low score means that detector found little evidence; it does not establish camera origin.
Google documents SynthID watermark checks in Gemini for supported images, video and audio. A detected watermark is evidence within that system’s scope. No detected watermark does not rule out another generator. Ask whether the result came from a watermark check or a visual guess.
Worked example: a marketplace product photo
Illustrative example: a screenshot has no metadata, a detector reports an uncertain result and the seller says it shows the item for sale. The correct origin conclusion is “unresolved.” Missing metadata does not make it generated.
For the purchase decision, ask for a fresh photo showing a specific angle and a handwritten reference, compare it with the listing and consider whether the seller can answer product-specific questions. This can add context but is not conclusive proof of ownership.
How to read our result
Our reports combine the findings available for the submitted file. A provider may be unavailable, and marker detection is not equivalent to cryptographic validation. Check which stages completed before interpreting a label.
We have not published a representative accuracy benchmark for this service. Do not use a detector result alone to accuse someone of deception. Preserve conflicting evidence and ask for additional sources.
Sources and further reading
Worked examples and scenarios are explicitly illustrative. Conditions and procedures can change; check the relevant original source.
Report an error or suggest a source