How to tell if an image is AI-generated
You cannot tell for certain whether an image is AI-generated from a glance, but four checks combined get close: look for visual artefacts, read the file's metadata and content credentials, reverse-search the image, and run it through a detector such as AiPicDetect. Each check can be fooled; together they rarely all fail at once.
1. Look for visual tells
- Hands, teeth and ears with the wrong count, merged fingers or asymmetric earrings.
- Text and logos that read as almost-letters, warped signage, gibberish on packaging.
- Lighting and reflections that disagree: shadows from different directions, mirrors showing a different scene.
- Repeating textures in hair, fabric, foliage or crowds, and backgrounds that melt into each other.
- Too-perfect skin and symmetry, especially with a shallow-depth-of-field look on every subject.
Current generators fix more of these every release, so the absence of tells is weak evidence.
2. Read the metadata and content credentials
Upload the file to AiPicDetect and open the metadata panel. Cameras write EXIF (make, model, exposure); editors write XMP; some generators write an XMP or IPTC tag naming themselves, and a growing number sign a C2PA content credential that records the generator. A signed credential is the strongest evidence you can get. No metadata at all is common (social platforms strip it) and proves nothing.
3. Reverse-search the image
Search the image with Google Lens, TinEye or Bing Visual Search. If the earliest copy is on a stock-photo site or a news archive from years ago, it is not freshly generated. If every copy traces back to one social account with no original source, be suspicious.
4. Run a detector and read the score correctly
AiPicDetect scores the pixels with an open-source classifier and returns an AI likelihood with a confidence band. A High-confidence result is a strong signal; Medium is a hint; Uncertain (within 10 points of 50%) means the pixels carry no usable evidence. Detectors are weakest on screenshots, compressed images and generators newer than their training data, as explained on the accuracy page. Try it on the detector, or self-host for private material.
Putting it together
| Evidence | Weight |
|---|---|
| Valid C2PA credential naming a generator | Conclusive |
| Generator name in XMP/IPTC | Strong (easy to strip, hard to fake by accident) |
| Reverse search finds an older real source | Strong for “real” |
| High-confidence detector score | Strong signal, not proof |
| Visual tells | Supporting |
| Medium or Uncertain score, no metadata | Inconclusive |
Laatst bijgewerkt 2026-09-13 · broncode op GitHub