Does Midjourney mark its images as AI-generated?

Midjourney is one of the most widely used AI image generators, historically accessed through Discord and more recently its own web app, and it's a common source of images for marketing, blogs, and social content. Unlike OpenAI and Adobe, Midjourney has been notably slower and less transparent about adopting machine-readable provenance marking, which matters directly for Article 50(2), since a business publishing Midjourney images can't assume the same default protection a Firefly or DALL·E user gets.

Article 50(2) of the EU AI Act requires AI-generated or AI-edited audio, image, video, and text to be marked in a machine-readable format so it's detectable as artificially generated, via C2PA Content Credentials or IPTC provenance metadata, embedded in the file itself. This is a provider-side duty on the generative system, but it fails quietly in practice: content that leaves a generator marked routinely arrives on a published page unmarked, stripped somewhere in the ordinary pipeline of editing, compressing, and publishing.

Does Midjourney mark its output by default?

As of August 2026, current reporting is consistent that Midjourney does not embed C2PA Content Credentials or any documented imperceptible watermark in its output — isolated 2026 roundup claims of C2PA adoption are not supported by Midjourney's own documentation or recent coverage. Midjourney has been a Content Authenticity Initiative member since 2023 but has not shipped an implementation. What IS confirmed: images downloaded from Midjourney's web interface carry ordinary XMP metadata (a dc:creator field with your account username, a dc:description field containing your full prompt, and an xmp:CreatorTool field identifying Midjourney), and Midjourney also sets the IPTC DigitalSourceType field to trainedAlgorithmicMedia and embeds the generation Job ID as a Digital Image GUID. That IPTC flag is a standardized, machine-readable ‘AI-generated’ signal, but it is unsigned and trivially strippable, not the cryptographically-signed C2PA manifest that Article 50(2) marking is built around.

What strips the marking before it reaches your published page

• Even the XMP metadata Midjourney does embed is ordinary file metadata, not a tamper-evident signed manifest, it's removed by nearly any re-export, screenshot, or basic image editor
• Discord itself re-compresses and re-hosts images, which can strip metadata as part of that processing before you even download the file
• Social media re-upload strips remaining metadata as a matter of course, the same as with any generator
• CDN-level image optimization on your own site removes whatever metadata survived to that point

How to verify before you publish

• Don't assume C2PA protection, check any specific Midjourney image with Content Credentials Verify (contentcredentials.org/verify); if no manifest is found, treat the image as unmarked for Article 50(2) purposes
• If a manifest isn't present, Article 50(4)'s separate requirement, that AI-generated content shown to the public be labeled, still applies; add a visible, honest label (e.g. "AI-generated image") near the image itself rather than relying on invisible marking
• Keep your original prompt and generation record as informal internal provenance evidence, even though it isn't machine-readable in the Article 50(2) sense
• Where machine-readable marking genuinely matters for your use case, consider running the finished image through a C2PA-aware tool (e.g. an Adobe Content Credentials-supporting editor) to attach a manifest yourself before publishing
• Re-check Midjourney's official documentation periodically, this is an area where the company's practices could change without prominent announcement

The realistic compliance gap

Midjourney is a genuinely higher-risk generator for Article 50(2) purposes than C2PA-native tools like Adobe Firefly or OpenAI's image products, precisely because its machine-readable marking status is unclear or unconfirmed rather than clearly present. A business relying on Midjourney for published visuals should default to the visible-labeling path (Article 50(4)) rather than assuming invisible machine-readable marking has them covered.

Common questions