How to Tell Whether an Image Was Generated by AI

The advice that circulated when image generators first became widely available was about artefacts: count the fingers, look for garbled text on signs, check for jewellery that merges into skin. That advice was useful and is now mostly obsolete. Current models render hands correctly most of the time and handle short text passably, and the specific failures people learned to look for are the exact ones each generation has been trained to fix.

Some visual signals still have life in them, but they are softer. Backgrounds remain the weakest area, because attention goes to the subject: look at what happens to architecture behind a crowd, at repeated faces in the distance, at patterns that continue where they should be interrupted. Physical consistency is another: reflections that do not match the scene, shadows falling in several directions, or a light source that does not correspond to anything visible. Overly even skin, symmetrical catchlights in both eyes, and a certain glossy uniformity across the whole frame are weak indicators rather than proof. Treat all of these as reasons to check further, never as conclusions.

The reliable approach is provenance rather than inspection. Ask where the picture came from, and specifically where it appeared first. A reverse image search on Google Images, TinEye or Bing will often surface earlier copies, and the earliest instance frequently answers the question on its own, either by leading to a photographer and a news story or by leading to a post that presented it as generated art before someone else reposted it as real.

Context is part of provenance. A dramatic photograph of a major event with no accompanying coverage from any outlet that would certainly have covered it is doing something odd, whether or not it was generated. Real events produce multiple photographers, multiple angles and reporting; a single image with no siblings deserves suspicion for ordinary journalistic reasons rather than technical ones.

There is also a growing technical layer worth knowing about. C2PA content credentials embed a signed record of how an image was created and edited, and support is arriving in cameras, editing software and some generators. Where credentials exist, they are strong evidence. Their absence, however, proves nothing at all, because the metadata is stripped by almost every social platform on upload, and by screenshotting.

That leaves detector tools, which invite more confidence than they deserve. They return a probability from a model that was trained on particular generators, and they misfire in both directions: real photographs that have been heavily edited or compressed get flagged, and output from a generator the detector has not seen gets missed. They are one input among several, not an adjudicator.

The honest summary is that visual inspection alone is a losing position and will keep getting worse, while the journalistic questions get more valuable as it does. Who published this first, what else did they publish, does any independent source show the same thing, and does the image behave like a real photograph of a real event with other witnesses. Those questions worked before generated images existed, and they are what continues to work now.