The surest way to tell if a photo was edited with AI is to find the unedited original and compare the two. When you can’t, check the file for edit labels (Content Credentials, or the note Google Photos adds to AI edits), then look closely at the places where an edit meets the rest of the picture: edges, shadows, reflections and texture. An AI detector helps with big changes like face swaps, but a small AI touch-up inside an otherwise real photo is the hardest case for any tool.
That last point matters. A fully generated image is AI from corner to corner. An edited photo is mostly camera pixels with a patch of generated ones, so the evidence is local, and you have to know where to look.
What counts as an AI edit? #
“Edited with AI” covers a wide range, from harmless to deceptive:
| Edit | What it does | Common tools |
|---|---|---|
| Object removal | Erases people, wires or objects and fills the gap | Magic Eraser, Clean Up, object erasers in phone galleries |
| Generative fill and expand | Adds new objects or extends the frame | Photoshop Generative Fill, Magic Editor |
| Moving things | Shifts a person or object to a new spot | Magic Editor |
| Face swap | Puts one person’s face on another’s body | Face-swap apps and web tools |
| Composites | Merges several shots into one “moment” | Best Take, Add Me |
| Enhancement | Sharpens, upscales or smooths skin | AI upscalers, beauty filters, Zoom Enhance |
The first four change what the photo shows. Enhancement usually changes how it looks. That distinction is worth keeping in mind, because a photo with an AI-smoothed face is edited, but it isn’t a fake in the way a swapped face is.
Step 1: Find the original #
An unedited copy is better evidence than any clue in the pixels. Run a reverse image search with Google Lens, TinEye or Bing Visual Search and look for versions with different framing, extra people, or an object that’s missing in the copy you have. News photos, event photos and product shots often exist in several versions.
If the photo came from someone you know, just ask for the original file. People who used an eraser tool to remove an ex from a holiday photo usually don’t mind saying so.
Our comparison of reverse image search and AI detectors explains when each one is the right tool.
Step 2: Check the file for edit labels #
Several editing tools now leave a note in the file:
- Google Photos notes when a photo has been edited with Google AI tools such as Magic Editor, Magic Eraser and Zoom Enhance, and shows it alongside the file name and location in the photo’s details. It uses IPTC metadata to do this, and also marks composites like Best Take and Add Me (Google).
- Content Credentials (the C2PA standard) can record which tools touched a file and whether AI was used. Adobe, Google, OpenAI and others are members of the group behind it (C2PA).
- Editing software tags in EXIF or XMP sometimes name the app that last saved the file.
A label is strong evidence of an edit. No label tells you little, because metadata is often stripped when a photo is uploaded, sent through a messaging app, or screenshotted. Our guide to checking image metadata for AI shows where to find these fields.
Step 3: Look where the edit meets the photo #
Open the largest copy you have and zoom in. AI edits tend to fail at the seams.
Edges and fills #
- A patch of background that is smoother or blurrier than the area around it.
- Repeated texture, such as the same bit of brick, grass or carpet appearing twice.
- Lines that bend or break where an object was removed: a railing, a horizon, a tiled floor.
- A faint halo or outline around a person who was moved or added.
Light and shadow #
- An added object with no shadow, or a shadow that points the wrong way.
- A person lit from the left standing in a scene lit from the right.
- Reflections in windows, mirrors, sunglasses or water that still show what was removed.
Faces and bodies #
Face swaps have their own tells:
- A color or texture change along the jaw, hairline or neck.
- A face that is sharper or smoother than the hair and ears around it.
- Skin tone on the face that doesn’t match the hands or arms.
- Glasses, earrings or teeth that look slightly melted.
Grain and noise #
Camera photos have fine, even grain. Generated patches are often cleaner. In a dim indoor photo, a region with no grain at all is suspicious.
Step 4: Run an AI detector, twice #
Detectors judge the whole image they’re given. If most of the pixels came from a camera, a small generated patch can be outvoted, and the result may come back “Likely real” or “Uncertain” even though something was changed.
A practical way around that is to run the check twice: once on the full photo, then again on a crop around the part you suspect, keeping the crop reasonably large so the model still has detail to read.
Expose AI is built for this on a phone. It targets both AI-generated and AI-edited images, including face swaps, and its Crop tool lets you isolate the region in question before you tap Analyze. Every result shows a verdict, a confidence meter and a “How we decided” card. If the file still carries Content Credentials or editor metadata, that shows up on the card as evidence, while the on-device model reads the pixels. Nothing is uploaded, which matters when the photo is of a real person.
Treat the result as one signal. A “Likely AI-generated” verdict on a crop around someone’s face, plus a visible seam at the jawline, is a strong combination. “Likely real” on a photo you suspect of a small eraser edit is not a clean bill of health.
What about error level analysis? #
Error level analysis (ELA), offered by sites like FotoForensics, highlights areas of a JPEG that were compressed differently from the rest. An edited region can stand out because it was saved fewer times. It’s a useful extra lens for people who know how to read it, but high-contrast edges and textures light up naturally, and a photo that has been re-saved many times flattens everything out. Don’t treat a bright patch in ELA as proof.
When an AI edit is a problem #
Most AI edits are ordinary: removing a stranger from a beach photo, cleaning up a product shot, smoothing skin. The ones that deserve scrutiny change what a photo claims to show:
- A news or protest photo with people added or removed.
- A marketplace photo with scratches or damage erased (see AI photos in marketplace listings).
- A dating profile with a swapped face or a heavily reshaped body.
- An insurance, refund or damage claim photo with damage added.
For those, keep asking for more evidence: a second angle, a different photo from the same moment, or a live video call.
Frequently asked questions #
Can an AI detector tell if a photo was Photoshopped? #
Sometimes. Detectors are trained to spot generated pixels, so large generative edits and face swaps often register. Traditional manual edits, like cloning or color changes, aren’t what these models look for, and small edits of either kind can be missed.
Does Magic Eraser leave a trace? #
In Google Photos, edits made with Google AI tools such as Magic Eraser are noted in the photo’s details through IPTC metadata. That note travels with the file only as long as the metadata survives, so a screenshot or a re-upload to a platform that strips metadata removes it.
Is a beauty filter an AI edit? #
Many modern beauty filters use AI models to smooth skin or reshape features, so yes, in a technical sense. Heavy filtering can push a detector toward “AI” even when the person is real, which is one reason detector results need context. See why real photos get flagged as AI.
How can I prove a photo was edited? #
The strongest proof is the unedited original, found through reverse image search or by asking the source. Content Credentials or editing metadata that name an AI tool are also strong. Visual clues and detector results support a conclusion but rarely prove it on their own.