Accuracy report · November 2026

How We Test Phrasly's AI Image Detector

How accurate is Phrasly's AI Image Detector? We test it on AI images from ChatGPT, Gemini, Midjourney, DALL-E, on real camera and phone photos, and on screenshots, compressed and re-uploaded copies. Here's how we test our AI image detector and what the results show.

Tested November 2026 · Production detector

Results at a glance

How Accurate Is Phrasly's AI Image Detector?

It caught every AI image in the test, from all 7 generators, and stayed just as accurate when the images were screenshotted, compressed or re-uploaded to social media.

Accuracy on AI images and real photos, as original files.
98.9%
Of AI images caught, from ChatGPT, Gemini, Nano Banana, Midjourney, FLUX, DALL-E and Stable Diffusion.
100%
Accuracy on screenshots, and 98.9% on social media re-uploads and compressed copies.
99.1%
Of verdicts correct when the image carried a verified AI signature (C2PA or SynthID).
100%

Examples

See Real Results From the Test

Three images from the test pack, with the verdict the detector gave each one.

AI-generated photo of an older fisherman mending a net at a harbour, made with ChatGPT

AI-generated · 100%

ChatGPT image, as downloaded

The file carried a verified AI signature (C2PA), so the detector could name the source: made with OpenAI.

AI-generated photo of a busy cobbled street lined with café tables and flower boxes, made with Nano Banana

AI-generated · 99%

Nano Banana image, metadata removed

We stripped all metadata and signatures and re-saved the file. The pixel scan still caught it.

Real aerial photo of snow-covered Mt. Hood taken from a plane on a Google Pixel 3

Likely real · 98%

Real phone photo

Mt. Hood from a plane, taken on a Google Pixel 3 in 2019. The detector called it real.

Photo: Kissmykumbaya, CC0, via Wikimedia Commons

Test set

What We Tested

We built a test set from the images people actually check: AI images from today's most-used generators, real photos from cameras and phones, and real photos with part of the picture changed by AI.

AI images

260 images from 7 generators: ChatGPT, Nano Banana 2.1, Gemini 3 Pro Image, Midjourney, FLUX.1, DALL-E 3 and Stable Diffusion 3. The ChatGPT, Nano Banana and Gemini images were made for this test in November 2026, so the detector had never seen them.

Real photos

184 freely licensed photos taken in 2020 or earlier, before today's image generators: 124 camera photos and 60 phone photos from Samsung, Google Pixel and Huawei phones. No user uploads: the detector doesn't store them.

Partly AI-edited photos

35 real photos where one area was replaced with an AI-generated object. We report these separately.

Image types

Landscapes, food, architecture, animals, faces, phone snapshots and digital art.

Real-world copies

Images rarely reach you as the original file. We scanned every image again in 5 altered versions, for 2,874 scans in all.

Method

How We Ran the Test

  1. Collect real photos from freely licensed public sources, all taken before today's AI image generators.
  2. Collect AI images from 7 generators: new ones made with ChatGPT, Nano Banana and Gemini for this test, plus public research sets for the others.
  3. Make real-world copies of every image: a screenshot, a social media re-upload, a half-size copy, a heavily compressed copy and a copy with all metadata removed.
  4. Scan every file once with the same AI Image Detector you use on this site, in November 2026.
  5. Score the verdicts. The detector answers AI-generated, Partially AI or Likely real. We count an AI verdict on a real photo as a false positive, and a Likely real verdict on an AI image as a miss.

How it works

Two Checks on Every Image

1. AI signature check

Many AI tools now sign their images with C2PA content credentials or a SynthID watermark. When a verified signature is present, the verdict is certain and names the generator. In our test, 21% of images were decided this way, and every one of those verdicts was correct.

2. Pixel scan

Signatures are easy to strip, so every image also goes through our image model, which looks for the patterns AI generators leave in the pixels. This is what caught AI images after their metadata was removed or they were screenshotted.

Results

Results by Generator and Real-World Copy

On original files, the detector caught 100% of images from every generator. Accuracy stayed close to 99% when images were screenshotted, re-uploaded, compressed or stripped of metadata.

Accuracy by type of copy

Screenshots99.1%Original files98.9%Social re-uploads98.9%Heavily compressed98.9%Metadata removed98.6%0%50%100%
AI images caught, by generator
GeneratorOriginal filesAltered copies
ChatGPT100%100%
Midjourney100%100%
FLUX.1100%100%
DALL-E 3100%100%
Stable Diffusion 3100%100%
Nano Banana 2.1100%90–97.5%
Gemini 3 Pro Image100%90–97.5%

Real photos correctly called real: 97.3% of original files, and 97.8–100% of altered copies.

Best results

When to Double-Check a Result

No AI image detector is right every time. In these cases, look closer before you decide.

A small AI edit inside a real photo

The detector judges the image as a whole, so a photo where only one small object was added by AI can still read as real. If you suspect a local edit, check the area closely and ask for the original file.

Large, high-resolution camera photos

The few real photos flagged in our test were large camera files, mostly marked "Partially AI". Treat a Partially AI result on a high-resolution photo as a prompt to look closer.

Google AI images that were stripped and re-shared

Most are still caught, but after the signature is removed and the image is compressed or re-uploaded, a few can slip through. Scan the earliest copy you can find.

GIF, HEIC and AVIF files

The detector accepts them, but this test didn't include them yet. For the most reliable result, convert to JPEG or PNG first.

Using your result

How to Read Your AI Image Detector Result

A result tells you how likely it is that an image was made or changed with AI. Before you act on it:

  • If the result names a generator, the image carried a verified AI signature. That verdict is certain.
  • Partially AI means part of the image looks AI-generated. Look closely at the details and the source.
  • Scan the original file where you can. Every copy, screenshot and re-upload removes some of the evidence.
  • Check where the image came from: who posted it first, and whether other photos of the same moment exist.

Want to check an image now? Try the free AI Image Detector.

What We Keep Private

We publish the results and how we got them. We don't publish the full test set, our scoring thresholds or how the image model is trained, because that would make the detector easier to fool.

How Often We Retest

We will rerun this test when a major new image generator launches, and update this page with the date and results each time.

FAQ

Frequently Asked Questions

Yes. In our November 2026 test of 479 images, it gave the right verdict on 98.9% of AI images and real photos. It caught 100% of the AI images, from all 7 generators we tested.

Check an Image for AI

Free, with signature checks and a pixel scan on every image.

Try the AI Image Detector