Detectors return confidence scores, and confidence scores are not evidence. What settles the question is provenance: who made the file, with what camera, and what record travelled with it. In between, there are visual tells worth knowing.
The visual tells worth checking first
- Text anywhere in the frame: signage, labels, packaging. Generators still mangle letterforms, spacing and kerning.
- Hands, ears and teeth, especially where fingers overlap or jewellery meets skin.
- Reflections and shadows that do not match the light source, or a window that reflects a different room.
- Repeating textures in the background: railings, tiles, crowd faces, books on a shelf.
- Hair edges against bright light, where strands fuse into a flat silhouette.
- Straps, glasses arms, earring backs: small physical objects that should continue behind something but do not.
What a real photograph has that a generated one often lacks
Photographs contain the noise of a real sensor, slight optical imperfections, and depth of field that behaves consistently. They also contain boring specificity: a scuff on a chair, a badly parked car, a sign in the wrong language for the street. Generated images tend toward a flattering average of the prompt, and the average is what the eye notices as too clean.
Metadata and provenance beats guesswork
Many cameras record capture data in EXIF, and an increasing number of editing tools and publishers attach Content Credentials, which are C2PA manifests describing what was captured and what was edited afterwards. If a file carries verifiable Content Credentials, you can check them rather than judging the pixels. If it carries nothing at all, that is not proof of anything either: social platforms strip metadata routinely on upload.
A quick sequence that works: ask for the original file when it matters, check its metadata, run the image through a reverse image search to find the earliest source, and prefer images whose origin you can name.
Why detectors are not a decision
Detection tools can flag a heavily edited photograph and can miss a careful generation. They are useful for triage inside a large library and unreliable as a final answer, particularly when the cost of being wrong is a brand claim about a person or a place.
The reliable answer is procedural: source images only from libraries of human photography, keep the photographer and licence with the file, and record what you used and where. That is what Photothentic does by design, which is a stronger guarantee than any confidence score.