What the documentation supports
Midjourney documents an Editor that works with its own generated images and with external images, which means a Midjourney-edited file can contain photographed material. We have not confirmed a complete official Midjourney provenance policy for SynthID or C2PA Content Credentials, so this page treats that coverage as unverified rather than universally absent. Check Midjourney's current documentation before assuming any particular signal is present.
Source: Midjourney Editor documentation (opens in a new tab)
Avoid visual stereotypes
Smooth skin, dramatic lighting or an unusually symmetrical composition can appear in both generated artwork and ordinary photography. Conversely, a generated scene may contain detailed imperfections. Counting fingers is not a reliable stand-alone test. Use visible inconsistencies as questions for the source: are reflections consistent, do signs make sense in context, and does the claimed event have corroborating records? These observations should remain tentative rather than turning into a verdict.
Ask which process was used
A useful provenance request names the specific image and asks whether it was generated, edited or photographed. If the creator chooses to share a creation record, such as a link to the job in their own Midjourney gallery or the original export, that context is usually more informative than a screenshot of a detector score. Absence from a public gallery does not prove camera capture. Image Evidence does not access private accounts, Midjourney generation histories or anyone's identity.
Read detector limits
Detectors learn from examples, and research shows that a classifier trained on one family of generators can fail on newer ones: Ojha, Li and Lee found that detectors trained on GAN images misclassified many images from diffusion models as real. A Midjourney release that was absent from a detector's test data is a blind spot until it is evaluated. A useful benchmark covers the relevant Midjourney versions, exports, edits and real photographs. Image Evidence has not published Midjourney-specific accuracy.
Follow up proportionately
For a casual post, the creator's disclosure may be enough. For a disputed image, ask for the original and check the surrounding claim against independent reporting. Respect ownership when sharing or uploading files. Image Evidence does not train on user content and provides no watermark removal, detector evasion, Humanizer, face recognition or image-to-person search. An uncertain result is a useful boundary, not a failure that should be hidden.
Can AI detectors detect Midjourney images?
Some detectors flag many Midjourney images, but no detector is reliable on every version, style and edit. Results drop when a model meets images unlike its training data, or when files have been resized, compressed or screenshotted. Treat any score as one clue, and look for published test results that include the Midjourney versions and file types you care about.
Does Midjourney add metadata or a watermark to images?
We have not confirmed a current, complete Midjourney policy for C2PA Content Credentials or invisible watermarks, so do not assume either is present or absent. Inspect the original download with a metadata viewer and check Midjourney's own documentation. Ordinary metadata can be removed or edited, so a missing field is not evidence that an image is a photograph.
How can I tell if an image is from Midjourney without a detector?
Ask the creator for the original export or a link to the generation record, compare the image with the claimed context, and look for earlier copies of the same image to find where it first appeared. Visual quirks can prompt questions but should not settle them. If none of that evidence is available, the honest answer may be that the origin is unknown.