Edited Images and How to Spot Them

A reader sends you a picture of a pickup truck half sunk in a flooded road, with the line "Millbrook Road after the culvert failure, today." It looks right. The water is murky, the sky is gray, a few people stand at the edge. Is it edited?

Maybe not at all. The three most common problems with that post have nothing to do with Photoshop: it could be from a flood two years ago, it could be from a different town, or a tighter crop could be hiding a second photographer's flood barrier that shows the road was closed. Pixel-level forensics answers the narrow question "was this image altered?" Fact-checking answers the wider question "is this claim true?" You usually need the wider one.

Editing is not the same as deceiving

Photographers adjust exposure, straighten horizons, remove a stray bit of lint from a lens. A newsroom with a decent standard allows those, and bans changes that alter the content: adding or removing people or objects, moving things, combining frames. The line is this: would a viewer, knowing exactly what was done, feel they were told something false about the scene?

The five common manipulations

1. Cropping

Cropping is quiet and powerful. A photo of ten people holding signs becomes a photo of two angry faces. A photo of an officer helping someone up becomes a photo of an officer's arm on a person. The check is to find the larger version. A reverse image search (lesson 03) often turns up a wider frame, because the same photographer or agency published the original. If you can only find the cropped version, say so and say what you could not see.

You can estimate how much was cut. Suppose the original is 4000 by 3000 pixels and the version going around is 1500 by 1000. The original has 4000 x 3000 = 12,000,000 pixels. The crop has 1500 x 1000 = 1,500,000. That is 1,500,000 / 12,000,000 = 0.125, or 12.5 percent of the frame. Roughly seven-eighths of what the camera saw is missing. That fact alone does not prove deception (a close crop is normal), but it tells you to look for the rest. The aspect ratios matter too: 4000 by 3000 is 4:3, and 1500 by 1000 is 3:2, so the crop was not just a resize.

2. Splicing

Splicing joins pieces of different images. Clues, in rough order of usefulness:

• Light direction. Shadows should point away from the light, and highlights on faces and objects should agree. A person lit from the left standing beside a car lit from the right is worth a second look. Be careful, since stage lights, windows, flash and reflections make several light sources common in real photos.
• Edges. A cutout often has a faint halo, a too-sharp outline or a soft fringe, especially around hair, fingers and glass.
• Scale and perspective. People or objects that do not fit the camera angle or the floor line. Heads sized wrong against their neighbors.
• Noise and sharpness. Every camera leaves a grain pattern, and the grain should look alike across the frame. A patch that is cleaner, blurrier or grainier than the rest may come from a different image.
• Resolution mismatch. Part of the picture is crisp and part looks enlarged.
• Reflections and shadows that do not exist. A person pasted in front of a window may have no reflection in it. A car with no shadow on a sunny road.

3. Retouching

The clone stamp copies pixels from one area to cover another. Done well, it removes a trash can or a person. A giveaway is repetition: the same bush pattern, the same crack in the pavement, the same cloud shape appearing twice in a place where the real world would not repeat. Some forensic tools look for duplicated blocks automatically. They also flag harmless repetition such as brick walls, fences, lawns and crowds, so a hit means "look here," not "forged."

4. Swapped captions

This is the workhorse of false posts. The image is real and the words are false. A flood in one town becomes a flood in another. A fire from a prior year becomes today's fire. A training exercise becomes an attack. You cannot find it by studying pixels. You find it by tracing the image to its earliest appearance: reverse search, the photographer or agency credit, the archive copy, and the weather and news of the day it claims to show. Lesson 03 walks through that process, and lesson 04 covers the date information a file may carry and why it can mislead.

5. Stripped context

Screenshots, reposts and cropped versions strip the credit line, the date, the link and the original caption. After three reshares, nobody knows who shot it. The question "who is the earliest source?" often leads nowhere because the trail was cut. In that case, the honest answer is "unverified." Do not fill the gap with a guess.

Forensic filters: what they show and what they cannot

Free forensic tools exist. The InVID-WeVerify plugin, a free browser extension for Chrome and Firefox, includes a magnifier, a forensic tab built around filters developed in a research project, a metadata reader for JPEG files, and one-click reverse search across several engines (checked October 2026). Used with care, those tools help. Used as a lie detector, they mislead.

Error level analysis, honestly

A JPEG is lossy: every time it is saved, some detail is thrown away. Error level analysis (ELA) re-saves the picture at a known quality, then shows the difference between your file and the re-saved one. The idea is that a region added later has a different history and shows a different error level. In a clean, once-saved photo, error levels tend to look fairly even across similar textures.

In practice ELA is noisy. Sharp edges, high-contrast text, saturated colors and flat white areas often light up in untouched photos. Smoothing, resizing, social media recompression and screenshots flatten or scramble the pattern. A forger who knows the method can tune an edit to defeat it. A bright patch can mean "pasted," or "sharp edge," or "text," or "a platform recompressed this." A uniform result does not prove the photo is untouched. So read ELA the way a doctor reads a blood test: a number that tells you where to ask questions. It does not produce a diagnosis.

Other filters and checks

• Magnification. Zoom into edges, hairlines, text on signs, reflections in eyes or glass, license plates. Details you can read may be checkable against other sources.
• Noise and luminance views. These help you see whether graininess or brightness patterns are consistent. Interpretation takes practice.
• Metadata. Camera, software and dates, if present. Lesson 04 explains why this is a weak witness: platforms strip it, and anyone can edit it.
• Double compression. Some tools estimate whether part of a JPEG was saved a different number of times. Advanced and not conclusive.

AI-generated images and the tells that expire

Many AI image generators are diffusion models: they start from noise and refine it toward a picture that matches a text prompt. Early versions left plenty of clues. Hands with too many or too few fingers. Garbled text on signs and shirts. Teeth that merge. Jewelry that melts into skin. Backgrounds where lines and furniture slide into each other. Perfectly smooth skin and a slightly glossy, painted look.

Those tells were real. They also expire. Developers fix the flaws people publish, and each new generation gets cleaner hands, readable text and believable lighting. A list of tells you memorized last year is a weaker test this year and a poor test next year. The honest rules are:

• A visible flaw is a reason to ask more questions. It is not proof, since real photos have odd hands too, and people can add flaws on purpose.
• A clean-looking image proves nothing. Good fakes will look clean.
• AI-image detectors give a score that can be wrong in both directions. The Content Authenticity Initiative itself says such tools can be valuable in some cases and that their results can be unreliable (checked October 2026). Lesson 06 and the AI detection course at /journal_tools/issues/45-ai-detection/ go into detector limits.
• The better test is a process test: who first posted it, is there a photographer or a camera original, do other independent photos or videos from the same event exist, and does it carry content credentials (lesson 07)?

A caution on over-correction. Once people learn that fakes exist, they start calling real photos fake. Researchers sometimes call this the liar's dividend: if everything might be fake, an authentic image can be waved away. Your goal is to be able to say "verified," "debunked" or "unverified," and mean it.

Questions and answers

Can I tell if a photo was edited just by looking?

Sometimes. Obvious cutouts, mismatched lighting and cloned patches show up to a patient eye. Skilled edits and AI images can pass. The useful answer is that looking is a screening step that helps you decide where to dig, and sourcing settles the question.

Is a retouched photo always misleading?

No. Removing dust or a lens flare changes nothing about what happened. The test is whether the edit changes the meaning of the scene. Publishers that care about trust state their rules.

What is the quickest check for an old photo passed off as new?

A reverse image search (lesson 03). Look for the earliest date, then look at the largest copy and the original caption. If the picture also exists in an archive or a stock library with a date, that is strong evidence.

Why do filters like ELA give false alarms?

Because compression behaves differently in different parts of any image. Sharp edges, text and saturated colors respond to JPEG saving more strongly than smooth skies do. Resizing and recompression add more noise. A bright zone is where to start asking questions.

Does a screenshot of a photo prove anything?

It proves that someone took a screenshot. It strips the camera details, the original compression and the credit line. Use it as a lead, and try to find the source it came from.

If a photo has a visible watermark, is it real?

No. A visible logo can be copied or added by anyone. A watermark that is cryptographically checked is a different thing (lesson 07), and even that proves where a file came from, not that the scene is honest.

Can a real photo have a shadow that looks wrong?

Yes, often. Mixed light from windows, flash and lamps can produce several shadow directions. Reflective surfaces add more. Check the physical scene if you can: where were the lights?

What should I do if I think a photo is AI-generated?

Ask where it came from. Look for a first poster, a photographer, other angles of the event, and content credentials. If a platform allows it, report it using the platform's synthetic media options (lesson 10). Say "I could not verify this image" unless you hold real evidence. See also AI detection.

Hard questions

If forensic filters are this unreliable, why bother with them?

Because they sometimes point to a specific spot worth asking about, such as an area that was cloned over a license plate. They are inexpensive and free. What you should not do is let a filter replace sourcing, or treat a filter result as a headline.

Will it become impossible to tell real photos from fake ones?

Detecting by looking alone is getting harder, and research groups that study this say so. That is why the field is moving toward provenance (who made it and how it changed) instead of guessing from pixels. Lesson 07 covers content credentials, which help when present and prove little when absent.

Should a local news site run a photo it cannot fully verify?

Many newsrooms will not, or run it labeled as unverified with the reason stated. A reasonable policy is to hold images from unknown accounts on breaking events until a second independent source, the photographer or a local official confirms. The cost of a few hours' delay is usually lower than the cost of a correction.

What if the edit is real but the claim around it is true?

This happens: a true story illustrated with an altered or staged picture. The photo can still deceive, even when the text beside it is accurate. Say what is accurate, say what is altered, and do not use the altered image as evidence.

Isn't it enough to say "this looks fake"?

No. If you are wrong, you have called an authentic image false, which hurts the person who took it and weakens your standing the next time. Offer evidence, or say "unverified."