How to tell if a photo was edited with AI

What counts as an AI edit

An AI edit starts from a real photo and changes part of it. Phones, photo apps and chatbots now do this in seconds, and the most common kinds are:

• Removing something: a stranger in the background, a power line, a stain, a scratch, a logo. Google's Magic Eraser, Apple's Clean Up, Samsung's Object Eraser and Photoshop's Remove tool all do it.
• Adding or replacing something: generative fill, Magic Editor, Galaxy AI's Generative Edit, or a request to a chatbot such as "put a dent in this car door".
• Expanding the frame: the tool invents what lay beyond the edges of the original shot.
• Changing a face: a face swap, a new expression, retouching that reshapes features.
• Restyling the whole picture: "make it evening" or "turn this into a studio shot" can repaint every pixel while keeping the scene.

The last two blur the line. A photo restyled from top to bottom is closer to a fully generated image than to an edit, while a removed power line leaves almost every pixel from the camera untouched. Where a picture sits on that line decides what any detector can tell you about it.

Step 1: look for an edit label

Most phone makers and big photo apps now record AI edits. The label usually sits in the photo's information panel, not on the picture itself:

• Google Photos (Android and iPhone): open the photo, swipe up or tap the info button, and look for "Edited with Google AI" in the details. Photos taken on a Pixel 10 also carry Content Credentials, a signed record that Google Photos shows in the photo's About panel and updates when the photo is edited.
• Samsung Galaxy: pictures changed with Generative Edit get a small sparkle mark in a corner, and the edit is written into the file's details.
• iPhone: the Photos app marks pictures changed with Clean Up in the photo's info.
• Adobe apps and many AI tools: Content Credentials can record which tool changed the file and how. You can read them with our Content Credentials checker.

A label that is present is good evidence that an AI tool touched the file. A label that is missing proves nothing. A screenshot, a crop, a re-save or an upload to most social and messaging apps strips the details panel, and one crop removes a mark in the corner.

Step 2: find other versions of the photo

Edits are easiest to prove by comparison. If the same scene exists somewhere else without the new element, or with a different one, you have your answer, and no detector can beat that kind of evidence.

1. 1

   Run a reverse image search

   Use Google Lens, Bing Visual Search or TinEye on the whole picture, then on a crop of the part you doubt. Where the tool allows it, sort the results by date. Our reverse image search guide walks through it.

2. 2

   Find the earliest copy

   The first upload is usually the closest to the original. Put it next to the version in front of you and compare: what was added, removed or moved?

3. 3

   Look for other angles

   Events, products and homes are rarely photographed only once. A detail that appears in just one picture, from just one account, deserves a closer look.

4. 4

   Ask for the original, or a new photo

   If the photo was sent to you for a claim, a listing or a story, ask for the file from the camera roll, or a second photo taken now from another angle. Edits rarely survive a request for the original.

Step 3: check the details an edit has to fit into

An edit has to blend into a scene the AI didn't create, and that is where it can slip. Zoom in on the part you doubt and compare it with the rest of the picture.

One inconsistency is a reason to dig further, not a verdict. Real photos have odd shadows too, especially with flash, mixed indoor light or a wide-angle lens.

What Perspicax can and can't tell you

Perspicax combines proven computer vision techniques with TrueEye 2.3, our state-of-the-art detector. It examines every region of an image, down to the pixels, and answers in one of three bands:

It is built and measured on photo-like images that were generated from scratch. On those, in testing, it caught 98.6% of images from the newest OpenAI, Google and ByteDance models on average, and wrongly flagged fewer than 1 in 2,000 original camera photos.

Those figures are about fully generated images. They say nothing about edits. When most of a picture's pixels came from a camera, the answer can be likely real even though one thing was added or removed, or inconclusive when the evidence is mixed. Read the answer like this:

• Likely AI-generated: the picture as a whole looks generated. A heavy restyle can land here too, so for a photo you were told is real, take it seriously.
• Inconclusive: not enough evidence either way. Go back to the label, the earlier versions and the details.
• Likely real: the picture as a whole looks like a camera photo. It does not rule out a local edit.

Face swaps and other deepfake photos

A face swap puts one person's face on another person's body in a real photo. It is an edit, so the same limits apply: because most of its pixels came from a camera, Perspicax may answer likely real or inconclusive on a face-swapped photo.

• Compare the face with other photos of the same person: ears, hairline, teeth and moles are hard to match.
• Look where the face meets the head: the jawline, the edges of glasses, hair crossing the cheek.
• Check the skin: a face with a different grain, sharpness or light from the neck and hands.
• Ask where the photo first appeared, and whether it was shared to prove a point about the person in it.

When the whole picture was generated, face included, it is in Perspicax's scope. Our guide to AI profile pictures covers made-up faces.

When an edit matters

Most AI edits are harmless: a removed photobomber, a brighter sky. They matter when a photo is used as evidence of something:

• Claims and refunds: a dent added to a car, mould added to a meal. See AI photos in refund and insurance claims.
• Listings: damage removed from a product or a room, a view added to a window. See product photos and real estate photos.
• News: a crowd made bigger, a sign changed, a person added or removed. See how to verify a news photo.
• Receipts: an amount, a date or a merchant changed. See fake receipts made with AI.

In each case the strongest check sits outside the picture: the original file, a second photo taken now, a live video call, or a record the photo should match, such as an order, a delivery log or a bank statement.