Edited vs AI Generated Images: How to Tell Photoshop from AI
"Fake" is not one thing. A smoothed-out portrait, a photo with a generative-fill sky and a person who never existed are three different problems that leave three different kinds of evidence. This guide maps the spectrum and explains what an AI image detector reports for each.
An edited image is a real photograph altered in software such as Photoshop, so its traces are local: clone repeats, edge halos and mismatched grain in one region. An AI generated image was produced entirely by a model such as Midjourney, so its traces are global fingerprints across the frame. Generative fill sits between: real pixels with an inpainted region.
- What is the difference between an edited and an AI generated image?
- What are the four levels of image manipulation?
- How do you tell if an image is photoshopped?
- How do you tell if an image is fully AI generated?
- What does generative fill leave behind?
- What does an AI image detector report for each type?
- FAQ
- Conclusion
The most common support question we get is "why did my real photo get flagged?" The answer is almost always that the photo sits somewhere in the middle of this spectrum. Knowing where tells you how to read the result.
What is the difference between an edited and an AI generated image?
The core difference is where the pixels came from. An edit modifies a real photo in specific places. AI generation creates the whole image at once. That one fact decides which traces appear and which detection method works.
| Edited image (Photoshop) | AI-assisted edit (generative fill) | AI generated image | |
|---|---|---|---|
| Origin | Real photograph, then altered by a person | Real photograph with a model-generated region | Created from a prompt, no camera involved |
| What changed | Specific regions a human selected | One or more regions filled by a model | The entire image, all at once |
| Typical traces | Clone repeats, edge halos, mismatched grain, warped lines near a slimmed body | A patch whose noise and texture statistics differ from the rest | Melted textures, garbled text, disagreeing shadows, model-level fingerprints |
| Metadata | XMP CreatorTool naming Photoshop, Lightroom or similar | C2PA record with c2pa.opened plus an AI action, when the tool writes one | C2PA c2pa.created with digitalSourceType: algorithmicMedia, or a generator tag, when present |
| What our detector reports | Low AI score; the source check classifies the editing tool separately from generators | Inpainting probability contributes to the AI score; the source check reads the C2PA record if it survived | High AI score; the source check names the generator when metadata survives |
What are the four levels of image manipulation?
Most people picture two boxes, real and fake. The reality is a spectrum with four points, and the middle is where almost all the confusion lives.
Original photo
A camera captured light from a real scene. The pixels carry one consistent sensor-noise pattern and, often, EXIF data listing camera, lens and capture time.
Retouched or edited photo
A real photograph altered by a person: skin smoothed, a blemish cloned out, a sky swapped, a body slimmed. The base image is real. Specific regions were changed with tools like the clone stamp or healing brush.
AI-assisted edit
A real photograph where a model generated new pixels inside it. Adobe's Generative Fill lets you select a region and have Firefly invent its contents (Adobe Help Center). The result is part camera, part model. In detection terms this is inpainting.
Fully AI generated image
No camera was involved. Midjourney, DALL-E 3, Stable Diffusion or Flux produced every pixel from a prompt. There is no original photograph underneath.
Traditional photo forensics was built for level 2. AI detection was built for level 4. Level 3 is the hardest for both, which is why our model scores it separately.
How do you tell if an image is photoshopped?
Editing tells are local. You are hunting for one region that does not match the rest of the photo.
Clone-stamp repeats. The clone tool copies one patch over another, so the same blade of grass, cloud or brick appears twice. Scan textured areas for identical twins.
Edge halos and soft cutouts. A pasted object often shows a faint glow, a too-clean edge or a colour fringe against its new background.
Mismatched grain. Every sensor leaves a characteristic noise pattern. A region inserted from a different photo carries different noise, which forensic tools can surface.
Warped backgrounds near a slimmed body. Liquify bends a waistline, and it also bends the door frame, tile grout or railing behind it. Straight lines that wobble near a person are a giveaway.
Inconsistent shadows on composited objects. A pasted object lit from a different angle casts a shadow that points the wrong way, or none at all.
Metadata naming editing software. XMP CreatorTool may read "Adobe Photoshop" or "Adobe Lightroom". That does not prove deception, since a crop counts too. It tells you the file was processed. Our source check classifies Photoshop, Lightroom, GIMP, Canva and Affinity as editing tools, separately from AI generators, so a Photoshop tag never reads as "made by AI".
One caveat: a skilled retoucher can hide every one of these. Manual inspection catches sloppy edits, not careful ones.
How do you tell if an image is fully AI generated?
AI tells are global and concentrated in the things models still render badly. Text and logos come out as garbled lettering because the model paints the shape of text, not words. Shadows point in conflicting directions. Skin is waxy and poreless. Backgrounds melt or repeat the same element. Hands fail in busy poses, though less often than they did in 2023.
That is the condensed version. For the full ten signs rated by reliability, see how to tell if an image is AI generated. For the fast yes-or-no version, use the 12-point checklist.
What does generative fill leave behind?
A generative-fill composite is a real photograph with an AI-generated patch inside it. Most of the pixels are authentic camera data. One region is pure model output. That breaks the assumptions both older methods rely on. Editing forensics expects the manipulation to be a copied or warped slice of real pixels, not invented ones. Whole-image AI detection expects generation fingerprints across the frame, but here only one region carries them, diluted by genuine photo.
Two things help. The first is a model that scores inpainting as its own class. Our image model returns separate probabilities for a fully generated image and for inpainted regions. We add the two to form the AI score, so an AI-filled patch pushes the score up even when most of the frame is real.
The second is provenance. C2PA, the Coalition for Content Provenance and Authenticity, defines a signed record attached to the file that lists the actions taken on it (C2PA specification). On 20 October 2026 I inspected the two public reference images published by the C2PA open-source project (contentauth/c2pa-rs fixtures). The generated reference carries a c2pa.created action and digitalSourceType: algorithmicMedia. The edited reference carries c2pa.opened, a c2pa.color_adjustments action and a c2pa.ingredient pointing at its parent image. The record itself tells you which level of the spectrum you are looking at.
The limit, which no detector escapes: re-saving the generated reference as a JPEG, resizing it, or re-encoding it as a PNG each removed the manifest entirely in the same test. Generative-fill evidence in metadata survives only until the first re-upload. After that, the inpainting score is what you have.
What does an AI image detector report for each type?
Our detector reads two layers and reports them separately. The pixel model returns an AI score from 0 to 100 built from generated-plus-inpainted probability. The source check reads EXIF, XMP, PNG text, IPTC and C2PA and returns a tool name with an evidence level of strong, moderate, light or none. Here is how each level of the spectrum tends to read.
A retouched real photo usually gets a low AI score. Smoothing skin or cloning out a sign does not create generation fingerprints. If the file still carries its XMP, the source check reports the editing tool as an editor, not a generator.
A fully AI generated image gets a high AI score, because the fingerprints are present across the whole frame. If the original file is uploaded, the source check often names the generator at strong or moderate evidence.
A generative-fill composite lands in between. The inpainting probability lifts the score in proportion to how much of the frame the model filled. A large filled region reads clearly. A small, carefully blended one may sit mid-range, and that mid-range score is the system being cautious, not broken.
The honest reasons a real photo scores high: aggressive AI upscaling, heavy denoising and extreme retouching can flatten a photo into the smooth, low-noise texture that generated images share. When the stakes are high, find a higher-resolution original and read the source check result before deciding.
FAQ
Can an AI detector tell if an image was photoshopped?
Only indirectly. An AI image detector targets generation and inpainting, not classic retouching, so a cloned-out blemish usually leaves a low AI score. The metadata layer can report that Photoshop or Lightroom touched the file. Detecting a specific manual edit is the job of image forensics tools that analyse noise and error levels.
Does generative fill count as AI generated?
Partly. Generative fill inserts model-made pixels into a real photo, so the result is a hybrid. A detector that scores inpainting separately will lift the AI score in proportion to the filled region. Our ranking of AI image detectors shows which tools score AI edits separately. The clearest signal is a C2PA record, which Firefly writes, listing an AI action, as long as the file has not been re-saved since.
Why did a real photo get flagged as AI?
Heavy retouching, AI upscaling or strong noise reduction can flatten a photo into the smooth, low-noise texture generated images share, which pushes the pixel score up. When this happens, look for a higher-resolution original, check the source check result for an editing tool rather than a generator, and treat a mid-range score as a prompt to look closer.
What is error level analysis?
Error level analysis, or ELA, re-saves a JPEG and maps how much each region changes. Areas edited after the last save compress differently and stand out. ELA is useful for spotting a pasted region in a photo. It says nothing about whether an image was generated by a model, and it fails on PNGs and heavily recompressed files.
Is photoshopping an image illegal?
Generally no. Retouching, compositing and generative fill are legal in most contexts. Legality turns on use: fraud, defamation, non-consensual intimate imagery and election deception are illegal in many jurisdictions regardless of the tool. See our deepfake laws by state overview for the US picture.
Conclusion: know which kind of fake you are looking at
Edited vs AI generated images is a spectrum, not a pair of boxes. Edits leave local traces and an editor's name in the metadata. Generation leaves global fingerprints and, when the file is original, a generator's name. Generative fill leaves an inpainted region that a two-class model can score. Read the pixel score and the provenance record as two separate pieces of evidence, and the result will make sense.
Scan the image with the AI image detector and see the generated and inpainted evidence for yourself. Free accounts include 50 detections per month. For the mechanics behind both layers, read how AI image detectors work.