How to check if an image is Photoshopped

Editing a photo leaves evidence in the pixels, whether or not the edit looks convincing. Finding that evidence is a different job from spotting an AI-generated image, and it needs different tools. This page is about the first one: cloning, compositing, retouching, and manipulation done by a person in Photoshop, GIMP or Affinity.

Error Level Analysis is the workhorse. Save a JPEG and every 8x8 block picks up a compression signature. Edit one region and re-save, and that region has now been compressed twice while everything around it has been compressed once. ELA re-saves the image at a known quality and subtracts the result, which turns that difference into something you can see: the edited area glows against a darker, evenly-compressed background.

Clone detection looks for the other common shortcut. The clone stamp and healing brush both work by copying pixels from somewhere else in the same image, so we search for regions that are statistically identical to other regions. A patch of sky duplicated to cover something is easy to spot this way, because real sky is never that repetitive.

Shadow physics catches composites. Every light source in a scene implies a direction, and every shadow in that scene has to agree with it. Paste a person from one photograph into another and their shadow almost never lines up with the shadows already there. People are famously bad at noticing this by eye, and it is trivial to check by geometry.

Frequency-domain analysis handles the rest. A camera lens is a physical low-pass filter, so a real photograph has a smooth frequency rolloff. Sharpening, upscaling and heavy retouching all disturb it in ways that show up in the Fourier spectrum long before they show up on screen.

None of these is conclusive alone, which is the point of running all of them. You get a per-signal breakdown rather than a verdict, so you can see which check fired and decide what it is worth.

Photoshopped or AI-generated? Two different questions

These are two different questions and they need two different answers.

A Photoshopped image starts life as a real photograph. Someone opened it and changed part of it: removed a person, smoothed a face, merged two shots, moved an object. Most of the pixels are still camera output. The forensic signals are local, and they are about inconsistency: this region was compressed a different number of times from that one, this shadow disagrees with that light source.

An AI-generated image was never a photograph. There is no camera, no lens, no sensor, and no original. The signals are global and they are about origin: the frequency signature of the model that produced it, the absence of sensor noise, EXIF that was never written because no hardware wrote it.

Which matters depends on what you are trying to establish. If you want to know whether a news photograph has been doctored, you want manipulation detection, and this is the right page. If you want to know whether an image was generated from a prompt, you want our AI image detector instead.

The awkward middle case is generative fill, where someone uses AI inside a real photograph to extend a background or remove an object. That is both things at once: a genuine photo with synthetic regions inside it. Our pipeline treats it as a manipulation with an AI signature attached, and the evidence chain will say so.

What this will not catch

Here is what will not work, so you can stop before wasting your time.

JPEG quality is the dominant variable in ELA, and it can defeat it. If a photo is edited and re-saved at very high quality, Q95 and above, the compression difference between edited and untouched regions can fall below what is measurable. Skilled editors know this. Some do it deliberately.

Screenshots of photos are close to worthless. Screenshotting flattens everything into a single new compression pass with a single uniform history, which erases exactly the inconsistency ELA looks for. If someone sends you a screenshot rather than the file, that is worth noticing on its own.

Social platforms are almost as bad. Facebook, Instagram, WhatsApp and X all re-compress on upload, sometimes more than once, and every pass moves the image further from its original compression state. Analysis still works, but confidence drops with each round trip. Ask for the original file where you can.

Small edits can hide under the noise floor. Removing a spot, whitening a tooth, nudging a horizon: these are legitimate and they are also below what pixel forensics can reliably separate from ordinary camera noise.

And the limitation that matters most: detection tells you an image was edited, never why. A magazine cover has been retouched. So has almost every product photograph, every headshot and most holiday snaps. Editing is not evidence of dishonesty. Whether an edit is deceptive depends entirely on context that no forensic tool can see.

When a doctored photo actually mattered

The Kate Middleton photograph in March 2024 is the case most people remember. Kensington Palace released an official family portrait for Mother's Day, and within hours Associated Press, Reuters, AFP and Getty had all issued kill notices withdrawing it, which is an extraordinary step for a palace handout. The tells were visible without any tooling: a misaligned sleeve, a hand that did not resolve, a zip that failed to line up. The Princess of Wales later said she had edited it herself. The interesting part is not the edit, which was trivial, but the reaction, which was not: four agencies decided that a photograph they could not vouch for was worse than no photograph.

The pattern repeats at lower profile constantly. Property listings with skies replaced and cracks removed. Dating profiles reshaped past recognition. Marketplace listings where the damage has been cloned out. Insurance claims with the wrong car. Court exhibits where a timestamp has been altered by a single digit.

What these share is that the edit is usually small and the consequence is not. Nobody fabricates an entire scene. They change one number, remove one object, smooth one detail, and rely on nobody looking closely enough to check.

Why professional edits are harder to catch

Professionals are harder to catch than amateurs, but the difference is smaller than you would think.

An amateur edit is loud. Someone opens a JPEG, uses the clone stamp on it, saves it back over itself at whatever quality the dialog defaults to, and ships it. That leaves an obvious double-compression boundary, often visible clone patterns, and metadata openly naming the software that did it.

A professional works differently. They edit non-destructively in a raw workflow, keep everything on layers, export once to a final file, and never re-save on top of an existing JPEG. That single export gives the whole image one uniform compression history, which is exactly what removes the ELA signal. Add deliberate grain to mask noise-floor differences and you have suppressed most of the pixel-level evidence on purpose.

What survives is physics. You can flatten compression history, but you cannot make a pasted shadow agree with a light source that was never there. Perspective has to be consistent across a composite, and reflections have to match the geometry of the scene. These take real effort to get right and are where careful composites usually fail.

Sensor noise is the other durable one. Every camera sensor has a unique noise fingerprint, and it is present across the whole frame of a genuine photograph. Content pasted in from a different camera carries a different fingerprint, or none at all. Uniform grain added over the top does not fix that, because the underlying inconsistency is still there beneath it.

In practice the interesting result is not a flagged image. It is an image where the compression evidence is clean but the physics is not, because that combination usually means someone competent has been at work.