How to Detect an AI-Generated Image

Begin with the story around the picture

Suppose an image claims to show a new hotel on Mars, a flooded street in Lisbon or a celebrity being arrested. Before enlarging anyone's hands, search for the event. If something that dramatic really happened, the picture should not exist in a vacuum.

Run a reverse-image search on the full image and then on a crop of the main subject. You may discover a perfectly real photograph from 2019 wearing a brand-new caption. That is misinformation, but it is not an AI-generated image. Keeping those two questions separate saves a great deal of confusion.

Follow the boring details

Faces attract attention, so they are often the strongest part of a generated picture. Background relationships can be less forgiving. Trace the arm of a pair of glasses. Follow a cable to its end. Count the legs under a crowded table. Check whether a repeated pattern survives when it bends around fabric.

Reflections are particularly useful. A shop window should agree with the street behind the camera; a mirror should not invent a different arrangement of furniture. The same goes for shadows and contact points. If a hand grips a cup, the fingers, rim and shadow should all describe the same physical action.

• Read small signs and labels from beginning to end.
• Compare earrings, buttons and clothing patterns on both sides of a person.
• Check whether objects at the same distance share a believable depth of field.
• Look for repeated textures that dissolve or merge at the edges.

Metadata helps, until it doesn't

An original file may carry camera details, editing software or a Content Credential. A valid credential can document the tool that created the picture and some of the edits made later. That is much stronger evidence than a hunch about skin texture.

Missing metadata proves almost nothing. Messaging apps strip it, websites resize files and screenshots discard it. Even an untouched photograph may arrive with an empty information panel. Treat provenance as a positive clue when it exists, not as a test every real image is required to pass.

Give the detector a fair file

Use the highest-resolution version you can obtain. Do not take another screenshot unless there is no alternative. Every round of cropping, saving and platform compression changes the texture and can make the analysis less reliable.

The result is one entry in your notes, not a rubber stamp. If it reports synthetic patterns, go back to the source and see whether the visual evidence agrees. If it reports no strong signal, say exactly that. 'This test found no strong signal' is honest; 'verified real' is a different claim altogether.

Evidence deserves better handling than a meme

For a harmless joke, you can stop after a few checks. If the image is part of an insurance claim, identity dispute, news report or court matter, keep the original rather than a downloaded preview. Record where it came from and what tools touched it. At that point a web detector is useful for triage, but it is not a forensic chain of custody.