Can you tell if an image is AI-generated?

Start with evidence, not a visual hunch

An image can look artificial because of editing, compression or style. An AI-generated image can look photographic. A visual impression alone is a weak basis for a claim about its origin. Keep the original file and where it came from if the question matters.

First, check for portable metadata: a generator field or Content Credential may provide useful context. This image-inspection guide shows how to inspect JPEG, PNG and WebP files with Free AI Wiper. If you find a credential, use a C2PA validator to check the signature; Free AI Wiper detects supported carriers but does not validate their signatures.

Second, consider whether the image may carry a vendor watermark. Google's SynthID is one example of a watermark embedded in generated media, rather than attached as ordinary metadata. A metadata report cannot answer whether that signal is present. A vendor's detector may be useful for its own outputs, but no single vendor test covers every image generator.

Third, examine the publication context: an original source, editing record, reliable witness or platform disclosure may help. Treat a generic classifier score as an uncertain lead. NIST's overview discusses the limits of automated content-based detection.

What if no signal is found?

You can say that a particular check found no supported signal in the file you tested. You cannot conclude that the image was made by a person. A screenshot or later export may have lost the original metadata, and another detection method may look for a different signal. See why metadata can disappear and what detectors can prove.