How to Tell if a Photo Is Fake

The most common fake is not an AI image at all. It is a real photo from another time or place with a new caption. Reverse image search finds the earliest copy and its original context in seconds.

Use Google Lens, TinEye or Bing Visual Search. If the same picture appears years earlier, under a different story, or as a stock image, you have your answer. If it appears nowhere, that is a mild signal it is new or generated, and you move to the next checks.

Cameras and phones write EXIF data into every photo: the device, the date and time, exposure settings and often GPS coordinates. Upload the original file to the EXIF viewer to read all of it, or to the metadata checker for a summary of what it says about the photo's origin.

What to look for: a camera model and capture settings support a real capture. A Software tag naming a generator, or Content Credentials declaring AI, settle it the other way. A date or location that contradicts the caption is a red flag. Missing metadata is not a red flag. Screenshots, messaging apps and every major social platform strip it, from real photos and fakes alike.

Photoshop, Lightroom, GIMP, Canva and phone editors leave records when they save: the Software tag, an XMP edit history, Adobe file markers, a file date later than the date taken, and Content Credentials actions like cropped or colour adjusted. The edited photo detector reads all of these and also flags AI-assisted edits such as Generative Fill.

A recorded edit does not mean the photo lies; most published photos are cropped or colour-corrected. It tells you the file was opened in an editor, which raises the bar for the next checks.

Once metadata is stripped, only the pixels are left. An AI image detector measures statistical patterns that generators leave behind and returns a probability that the image was AI-generated. The uploader at the top of this page runs it; so does the homepage.

Read it as evidence, not a verdict. A high score on a stripped image is a strong signal it was generated. A low score on an image with intact camera metadata is a strong signal it is real. A middling score means the detector is unsure and the other checks decide. For faces specifically, the deepfake detector is tuned for that case.

• One light source. Shadows in a real scene agree on direction and softness. An inserted person or object often casts a shadow that points the wrong way, or none at all.
• Reflections. Windows, water, glasses and eyes should reflect the same scene. Missing or mismatched reflections are a classic composite tell.
• Edges. Zoom in where a subject meets the background. Cut-and-paste edges are too sharp or have a faint halo; AI images blur where hair, fingers or glasses meet skin.
• Perspective. Lines that should converge (floor tiles, buildings, tables) do so consistently in a real photo and drift in a fake.

• Text. Signs, labels, logos and clothing text in AI images are often garbled or almost-words. Real photos have legible, consistent text.
• Hands and teeth. Generators still miscount fingers, merge them, and produce too many or too few teeth.
• Repetition. Repeated crowd faces, identical texture patches, or the same object twice point to cloning in an editor or to a generator's tiling.
• The sheen. A glossy, over-smooth, perfectly lit look across skin, fabric and background is characteristic of generated images.

Full list of visual tells: how to tell if an image is AI-generated.

Who posted it first, when, and do reputable outlets show the same scene from other angles? A genuine news photo of a real event is rarely a single image from a single anonymous account. Check the account's history, the timestamp against the event, and whether the caption is doing work the pixels do not support.

This is the check that catches the most fakes in practice, because most fakes are real photos with false claims attached. Fact-checkers and newsrooms run it alongside the tools above; see how they use wedetect.

No single check is proof. Two or three that agree are.