AI vs Real Images: 11 Examples Side by Side

Often not by eye. Below are 11 AI vs real images from our detector test: four pairs where one is AI-generated and one is a real photo, then three cases that gave the software trouble. Under each pair you'll find which picture is which, what the file said and how two detection models scored it; the table above sums them all up.

Every AI picture here was generated by us for the test; every real photo is a CC0 picture credited to its author. The pictures on this page are small copies with the metadata removed, so they won't give the same result as the originals if you check them. To test a picture of your own, use the AI image checker.

Both pictures are wide mountain views with water and soft light, the kind of scene generators handle very well. Look at the sky, the trees and the water, then read what the file and the detectors said about each one.

Which is which:

• A is a real photo. Camera metadata: Apple iPhone 11; no Content Credentials. Sightengine scored it 0.001, Hive 0.00. Our checker said Likely a real photo. Photo: Riddles42, CC0 1.0.
• B is AI-generated. Made by us with GPT Image 2; the file carried Content Credentials signed by OpenAI OpCo, LLC, a signer on the C2PA trust list. Sightengine scored it 0.99, Hive 1.00. Our checker said Likely AI-generated from the credentials alone, without calling a model. Prompt: "Wide landscape photo of a misty mountain valley at sunrise with a winding river and pine forest, shot on a full-frame camera"

Food photos sell dishes on menus and delivery apps, and generators are good at them. One of these was shot on a phone in an ice cream shop; the other came from a text prompt asking for an overhead phone photo of ramen.

Which is which:

• A is AI-generated. Made by us with Nano Banana 2 (Gemini 3.1 Flash Image); the file carried Content Credentials signed by Google LLC, a signer on the C2PA trust list. Sightengine scored it 0.99, Hive 1.00. Our checker said Likely AI-generated from the credentials alone, without calling a model. Prompt: "Overhead phone photo of a bowl of tonkotsu ramen on a restaurant table, steam rising, chopsticks resting on the bowl"
• B is a real photo. Camera metadata: Google Pixel 7a; no Content Credentials. Sightengine scored it 0.08, Hive 0.00. Our checker said Likely a real photo. Photo: Mx. Granger, CC0 1.0.

Product photos on a plain background are hard to judge by eye, because there is so little detail to go on. Both of these would pass as a shop listing. One is a stock photo; the other was made from a prompt.

Which is which:

• A is a real photo. A stock photo with no camera metadata; no Content Credentials. Sightengine scored it 0.001, Hive 0.00. Our checker said Likely a real photo. Photo: Icons8 team, CC0 1.0.
• B is AI-generated. Made by us with GPT Image 2; the file carried Content Credentials signed by OpenAI OpCo, LLC, a signer on the C2PA trust list. Sightengine scored it 0.99, Hive 1.00. Our checker said Likely AI-generated from the credentials alone, without calling a model. Prompt: "E-commerce product photo of a matte black stainless steel water bottle on a pure white background, soft studio lighting, subtle shadow"

Street scenes give you more to check: signs, reflections, people and lines that should stay straight. One of these pictures was taken on a phone on a wet winter day; the other was generated from a one-line prompt.

Which is which:

• A is AI-generated. Made by us with Seedream 4.5; the file carried Content Credentials signed by Byteplus Pte. Ltd., a signer not on the trust lists we check. Sightengine scored it 0.99, Hive 1.00. Our checker said Likely AI-generated. Prompt: "Handheld photo of a crowded night street market with food stalls and neon signs, light rain"
• B is a real photo. Camera metadata: samsung Galaxy S24; no Content Credentials. Sightengine scored it 0.001, Hive 0.00. Our checker said Likely a real photo. Photo: Estormiz, CC0 1.0.

Detectors make mistakes too, and the pictures that trip them up are worth seeing. These three are the hardest cases from our test: an AI landscape that one engine missed, a real product photo that ended in Can't tell, and an AI wedding photo that our checker called real once it had been screenshotted.

AI-generated (FLUX.2 [pro], Content Credentials signed by Black Forest Labs Inc., a signer not on the trust lists we check). Sightengine scored it 0.99, Hive 0.47. Our checker said Likely AI-generated. Hive's 0.47 is below its recommended 0.9, so on its own Hive would have missed it. Made by us with FLUX.2 [pro] for this test. Prompt: "Autumn forest path covered with fallen leaves, soft fog between the trees, photographed on a mirrorless camera"

A real photo, turned into a phone-style screenshot for the test. Sightengine scored it 0.21, Hive 0.11. Our checker said Can't tell. At 0.21 the score sat just above the 0.2 line for Likely a real photo, so the honest answer was Can't tell. Photo: W.carter, CC0 1.0.

AI-generated, turned into a phone-style screenshot for the test. The original file carried signed credentials; the screenshot carried none. Sightengine scored it 0.18, Hive 0.00. Our checker said Likely a real photo. This was the only AI image our checker called real in the whole 100-image test. Made by us with FLUX.2 [max] for this test. Prompt: "Candid wedding photo of a bride and groom dancing under string lights at an outdoor evening reception, guests blurred in the background"

Mostly the file, rather than anything we could see in these AI vs real images. In 3 of the 6 AI examples, signed Content Credentials from OpenAI or Google settled it before any model looked at the picture. For the others, the detection model did the work.

The one visual tell we found was in the Seedream night market: on the full-size file, the neon signs are made of shapes that look like Asian script but, as far as we can tell, don't spell real words. It is hard to see in the small copy here.

That is the honest lesson from these examples: newer generators leave fewer of the classic mistakes, and real photos can look just as polished. Our visual checklist still helps with signs, small text and crowded scenes, but treat what you see as a clue. Two generators in this set, FLUX and Seedream, sign their files with a signer that isn't on the trust lists, so the Seedream detector and FLUX detector pages explain why those still go to a model.

To sort your own AI vs real images, use the checker on this page with the original file whenever you can get it, because screenshots lose the credentials that settled most of these examples. It reads Content Credentials and metadata in your browser first, then asks Sightengine to judge the pixels, and tells you which signal it relied on.

Across all 100 images in the test, that flow judged 98 correctly, answered Can't tell for 1 and got 1 wrong, with none of the 50 real photos flagged as AI. The accuracy report has the full method, and the metadata checker shows what a file says about itself.

They come from our 100-image detector test, not from a search for tricky pictures: we paired images with the same kind of subject and left out real photos of people. We know which is which because we generated or sourced every file ourselves.

Putting a real person's face next to the question "real or AI?" invites people to doubt that person. The AI portraits in our test are fine to show; the real ones stay in the test set.

You can, but the copies here are resized and stripped of metadata, so they test only the pixels. A detector that relies on Content Credentials will give a different answer than it would on the original files.

OpenAI GPT Image 2, Google Nano Banana 2, ByteDance Seedream 4.5 and Black Forest Labs FLUX.2. The full test also covered Midjourney and Stable Diffusion XL.