Midjourney Detector

Midjourney didn't sign any of the 7 original files in our test with Content Credentials, so the credentials step never settles a Midjourney picture. Every one goes to a detection model that judges the pixels: Sightengine first, Hive if Sightengine fails. That uses one of your 3 free checks a day, unless someone has already checked the same file.

Add the picture to the Midjourney detector above and you get one of four answers, with a range and the reasons behind it. Your browser still reads the file's metadata first, because some Midjourney files say they are AI-generated in a field anyone can edit (more on that below). The flow is the same one our home page uses to detect AI images from any generator.

Sightengine called all 8 Midjourney images AI, every one at or above the checker's 0.8 line, with a lowest score of 0.99. Hive caught all 8 too, and the checker got 8 right out of 8. That is a strong result on a small sample: 7 original files and 1 screenshot, tested on October 13, 2026.

We generated 5 of the originals ourselves with Midjourney v7 and v8.1 and took 2 Midjourney v6 images from a public dataset. None of the 7 carried Content Credentials, whichever way we got them, and neither did the screenshot. Midjourney and Stable Diffusion XL were the only families in our test without any credentials; the accuracy report has the other four.

4 of the 7 Midjourney originals carried a DigitalSourceType field in their metadata with the IPTC value for Created using Generative AI. That is a useful hint, but it isn't signed: anyone can add it to a real photo, or delete it from an AI image, with a free metadata editor. So the checker treats it as the file's own claim and still asks the model.

What you see then depends on the model. If it also says AI, you get Likely AI-generated, with a note that the file says so too but the label can't be verified. If it scores the picture like a real photo, you get Can't tell, because the file and the model disagree. If the model isn't sure, you get Can't tell with both the label and the score listed. In our test the model agreed with all 4 labels. A label never decides on its own, which protects a real photo that someone has tagged as AI.

The checker can say a picture is likely AI-generated, but it usually can't say it came from Midjourney. With no credentials to name the tool, the answer rests on the pixels, and Sightengine, the main engine, returned only an overall score in our test.

Hive, the backup engine, does guess a generator. When Hive handles a check and calls the picture AI, the result can name the likely tools, with a note that this is a guess from the pixels. Treat that as a lead, not proof. A label in the file is no better, because it is as easy to change as any other text. If it matters which tool was used, look for the first post: people who share their Midjourney work often say so in the caption.

A Midjourney detector result is one piece of evidence about a picture, never a verdict on the person who posted it. Here is how to use each answer.

• Likely AI-generated. Strong in our test, where Sightengine missed no Midjourney image. It's still a probability, so quote the range.
• Can't tell. It can mean a heavily edited or compressed copy, or a file whose label and model disagree. Find a better copy before you decide.
• Likely a real photo. We never saw this on a Midjourney image, but 8 images can't rule it out. If the source seems off, keep going with the visual checklist.

For contest entries and artwork, read the AI art detector page before you act on any score. For a family that was much harder to catch, see FLUX.

None of the 7 Midjourney originals in our test had them, including the 5 we generated ourselves with v7 and v8.1. So the Midjourney detector always relies on a detection model, and a missing credential tells you nothing.

No. It answers whether the picture is likely AI-generated, not which model or version made it. We tested v6, v7 and v8.1, and Sightengine caught all of them.

Because there are no credentials to settle it in your browser, the picture has to go to a detection model. If someone already checked the exact same file, the stored result comes back without a new check.

Because the label is unsigned and the detection model disagreed or wasn't sure. The checker won't call a picture AI on an editable label alone, and it won't call it real while the file says otherwise.