AI ghost photos: how a generated image gives itself away

The photograph usually arrives the same way. Sepia, a staircase or a hallway, a pale figure on the landing, and a caption saying it turned up in an attic or a box of family prints. When an image like that was made by an AI model, it gives itself away in the parts nobody looks at: the hands, the lettering, the window behind the figure, the way the railing meets the wall. Image models paint convincing texture with no consistent geometry underneath, and the geometry is where you catch them.

What follows is where to look, in roughly the order that pays off, and what to ask of the file itself.

How an image model makes a photograph

Midjourney, Stable Diffusion, FLUX and DALL·E share a basic method. The model starts from an image of pure noise and removes that noise step by step, each step nudging the picture toward what the prompt describes. At the end there is a photograph-shaped image that no lens ever saw.

The result is plausible at every small patch. Wood looks like wood, skin like skin, a curtain like a curtain. What the model does not keep is a single, consistent world behind the patches: one light source, one set of parallel lines, one number of fingers. A camera cannot help recording a consistent world, because there was one in front of it. A model has to fake that consistency, and it often fails at the seams.

Look at the hands, the text and the small metal

These are the classic tells, and they still turn up.

• Hands. Count the fingers. Extra fingers, fused fingers, a thumb on the wrong side, a hand that dissolves into a sleeve. Look at the living people in the frame too, not only the figure.
• Text. A sign on the wall, a book spine, a date written on the back of the print. Generated text tends to turn into letter-shaped marks that look like writing from a distance and read as nothing up close.
• Jewelry and eyeglasses. Earrings that merge into the earlobe, a necklace that becomes part of the collar, glasses whose frames melt into the skin or do not match from one side of the face to the other.

Look at the room

This is where a careful fake still slips, because rooms are made of straight lines that have to agree.

Window panes should be the same size across a window and line up with each other. Railings should be evenly spaced and meet the floor. Floorboards run parallel and recede toward one vanishing point. In generated images they wander, split, change width halfway across the frame, or run into the wall at an angle the wall cannot have.

Then find the light. A real room has a window or a lamp, and every shadow falls away from it. Reflections in a mirror or a glass door show what is actually in front of them. In generated images shadows point in different directions, reflections show a different room, and some surfaces are lit by light with no source at all: a face glowing evenly in a dark corridor with nothing to glow from.

Two last textures. Patterns such as wallpaper, tiles or lace repeat faithfully for a while and then break, the motif changing shape without a seam. And the whole image can have a uniform smoothness, skin and plaster and fabric all polished to the same even finish, which a real lens and real grain rarely produce.

The fake age of a “vintage” ghost photo

AI ghost images are often prompted to look old, because an old photograph comes with a ready-made story. The prompt says something like “vintage ghost photo,” and the model supplies the look: grain, sepia, scratches, a vignette.

Look at how that damage is distributed. In a generated image it is spread evenly, grain at the same density everywhere, scratches crossing the figure and the background alike, the sepia a single tone laid over the whole frame like a filter.

A real print ages as an object. Damage follows the paper: creases where it was folded, worn and rounded corners, silvering or stains where chemicals and fingers reached it, fading that is stronger at the edges or where light fell on it in a frame. A glass plate cracks along lines and loses emulsion in patches. If the damage on the image does not follow the shape of a physical object, it was probably not done by time.

What the file says

If you can get the file itself, open its metadata. A photograph from a camera or phone usually carries EXIF data: the camera model, the lens, the exposure, often the date. A generated file often has no camera EXIF at all, or names only the editing software it passed through.

This is a signal, not a verdict. Many platforms strip metadata on upload, so a real photo that has been reposted may have none either, and metadata can be edited. What matters is whether the file’s story holds together. A photo described as a scan of a 1920s print should look like a scan, with the edges of the print and the texture of the paper.

Four checks to run

1. Reverse image search. Run the image, and a crop of the figure, through a reverse image search. You may find the first post, a version with a model’s watermark, or the account that made it.
2. Look for content credentials. Some generators attach content credentials under the C2PA standard, a signed record of how the image was made. When they survive, they answer the question directly. When they are absent, they answer nothing.
3. Ask for the original. Ask the person who posted it for the full-resolution file, or for a photo of the physical print, front and back. A real print is an object, with a back and a history of its own.
4. Compare against real aging. Put it next to a real plate or print from the period it claims. Look at the edges, the corners and the grain, not the figure.

If a photograph passes all four and you still cannot place it, send it to us. Moving images have their own set of tells, covered in how AI ghost videos are made, and some real photographs carry artifacts that look supernatural for ordinary reasons, set out in glitches that look like ghosts.

The older trick

None of this is new in spirit. The first ghost photographs, from the 1860s, were double exposures: William Mumler, a Boston engraver, found a faint figure on a reused glass plate in 1862, and the spirit photographs that followed were made the same way. The story of that plate opens the On Camera section.

The difference matters.

There was no one on the landing, and no landing.

Which leaves the photographs that are neither: a plate with no second exposure anyone can find, a print with honest creases and a figure no one in the family can name.