Practical guide

How to Tell If an Image Is AI-Generated

Short answer: no single visual artifact, metadata field or detector score can prove that an image is AI-generated. Start with provenance, inspect the original file, compare independent technical signals and verify the surrounding source. If the evidence conflicts or the original is unavailable, the defensible result is inconclusive.

A five-step checking process

  1. Preserve the original file.A screenshot or social-media download may remove metadata, recompress pixels and detach Content Credentials. Ask for the earliest available file before drawing conclusions.
  2. Check provenance first.Look for valid Content Credentials and assess whether the signer is trusted for the claim being made. Presence can add strong context; absence is neutral.
  3. Inspect file evidence.Review format, encoding, EXIF/XMP blocks and editing history where available. Metadata can be edited or stripped, so treat it as a claim that needs corroboration.
  4. Review visual patterns.Look for inconsistent geometry, repeated local texture, implausible reflections, text distortions or boundary errors. Modern generators may avoid these clues, and ordinary editing can create similar artifacts.
  5. Compare independent evidence.Seek agreement across provenance, file structure, visual analysis and source context. Record disagreements and stop at inconclusive when the evidence is insufficient.

What each signal can and cannot establish

SignalCan supportCannot establish alone
Valid, trusted Content CredentialA signed provenance claim is bound to the inspected asset.That every real image has credentials, or that the depicted event is true.
EXIF or XMP metadataCamera, software or workflow claims recorded in the file.Authenticity; metadata may be edited, removed or copied.
Visual anomalyA region worth closer review.The generator, author, intent or full edit history.
Detector scoreHow one model responded to one processed file.A universal probability that the image was made by AI.
No obvious signalNothing strong was found by the available checks.Authentic camera capture.

When should the answer stay inconclusive?

Keep the result inconclusive when only a screenshot is available, signals disagree, provenance cannot be validated, the detector’s evaluation scope is unknown, or the decision could materially affect a person. A cautious answer is more useful than a confident claim unsupported by the file.

For the model-output boundary, read how to interpret AI detector scores. For signed provenance, read Content Credentials explained.