Midjourney, DALL·E, and More: Labeling AI Images Correctly

Last updated: July 23, 2026

You have generated an image with Midjourney, DALL·E, Stable Diffusion, Adobe Firefly, or Nano Banana, and you are wondering whether the generator has already taken care of the legally required labeling itself? The short answer: partially, but you should not rely on it. This article shows what the common AI image generators actually write into the file, why that generally does not satisfy the obligation under Article 50 of the EU AI Act from August 2, 2026, and how you can properly complete a generated image in a few seconds.

Key takeaways

• The generator marks, if at all, only in a machine readable way. For the visible deployer labeling (Art. 50(4)), you are responsible, not the tool.
• The generators' metadata often does not survive the upload. Social platforms and marketplaces remove it when recompressing.
• Not a single generator places a visible label in the image. Yet that is exactly the level that must be recognizable to humans.
• Your role is that of the deployer. You use the AI system and publish the output, so you carry the disclosure obligation when the image is a deepfake, that is, looks deceptively real.
• Solved in one pass: visible EU icon, IPTC/XMP, optionally C2PA and watermark, plus proof, with the free tool from AI labeling.

Contents

1. Is the labeling the generator provides enough?
2. What the individual tools embed (overview)
3. Why you still need to complete it as the deployer
4. How to properly label a generated image
5. Edge cases: upscaling, inpainting, variants
6. Summary
7. Frequently asked questions (FAQ)

You can find a general step by step guide independent of the tool under labeling AI images, and the legal classification in the guide to the AI labeling obligation.

Is the labeling the generator provides enough?

In most cases: no. The reason lies in how the AI Act distributes roles. The provider of an image generator must mark synthetic outputs so that they are machine readably recognizable as AI (Art. 50(2)). Some providers satisfy this via C2PA Content Credentials or an IPTC field, others do not do so at all so far. The disclosure visible to humans, on the other hand, is the responsibility of the deployer, meaning you, when you publish a deceptively realistic image (a deepfake) (Art. 50(4)).

Even where a generator embeds machine readable data, this level is fragile. If you upload the image to Instagram, TikTok, or a marketplace, it gets recompressed, and the credentials frequently disappear. Not a single generator sets the visible label that no platform can compress away. That is exactly the gap you have to close.

Rule of thumb: at best, the generator handles a "light version of the provider level." The deployer obligation (visible label) always stays with you.

What the individual tools embed

The overview below summarizes the typical status. It reflects observed experience and can change with every model version, when in doubt, check the current file yourself.

Two patterns stand out. First, none of the tools set a label visible to humans, that always stays your job. Second, the machine readable level is inconsistent, ranging from a "solid C2PA chain" (Firefly, DALL·E) to "nothing at all" (local Stable Diffusion). How the IPTC, C2PA, and watermark formats work together technically is explained in C2PA, IPTC, and watermarks.

Why you still need to complete it as the deployer

Assume your generator delivers clean C2PA credentials. Then you are already well positioned at the machine readable level, as long as the file is not recompressed. Still, two obligations remain open:

• The visible disclosure (paragraph 4). For deepfakes and deceptively realistic images, a human must be able to recognize, without any tools, that AI was involved. Content Credentials are invisible, they do not satisfy this level.
• Robustness against platforms. As soon as Instagram, TikTok, or a marketplace touches the file, the machine readable level is at risk. The icon burned into the pixel image survives.

Whether you are even obligated as a deployer at all is covered in the article provider or deployer. For the vast majority of commercial users, shops, brands, agencies, the answer is: yes. And in the event of violations, fines of up to 15 million euros or 3 percent of worldwide annual turnover are possible (EU guidelines on the transparency obligations).

How to properly label a generated image

The process is the same regardless of which generator the image comes from:

1. Set the visible EU icon. Fully generated by prompt? Use the "fully AI generated" motif. Only partially AI altered (for example inpainting on a real photo)? Then "partially AI modified." Place it in a corner with good contrast so cropping does not cover it.
2. Embed or preserve the machine readable marking. Set the IPTC/XMP field DigitalSourceType and, where possible, C2PA. If the generator already provides credentials, preserve them by not re exporting in a lossy way.
3. Check after the upload. Download the published version back and use the check file function to verify that the icon and metadata were preserved.
4. Document proof. Store the SHA-256 hash and timestamp in the register, in case of a later inquiry.

With the free web tool from AI labeling, you handle steps 1 and 2 in a single pass: drag the image in, choose the motif and placement, export. The official EU icon is burned into the pixel image, IPTC/XMP (optionally C2PA and a watermark) get embedded, and everything runs locally in the browser, without your image being uploaded to a third party server. For several images from one prompt run, the batch function processes up to 50 images at once.

Edge cases: upscaling, inpainting, variants

• Upscaling an AI image: remains subject to labeling, the image is still AI generated.
• Inpainting on a real photo: "partially AI modified," because generative content is added.
• An AI variation of an AI image: still "fully AI generated."
• A real photo, only AI upscaled or denoised: generally uncritical, as long as no content is generatively added. The distinction is explained in detail in labeling AI images.

For deceptively realistic depictions of people or scenes, the stricter deepfake logic additionally applies, see deepfake labeling. You can look up unclear terms in the glossary.

Summary

• No image generator sets a visible label, this deployer obligation stays with you.
• The tools' machine readable marking is inconsistent and often does not survive uploads.
• So for content that requires labeling, complete it yourself: a visible label (the optional EU icon or a clear text note), set or preserve metadata, check after the upload, document proof.
• Whether Midjourney, DALL·E, Firefly, Stable Diffusion, or Nano Banana, the labeling step is the same and is done in seconds with the tool at ki-kennzeichnen.de.

Frequently asked questions (FAQ)

Does Midjourney label my images automatically?

No, not at the level that matters. Midjourney does not set a KI label visible to humans in the image, and a reliable machine readable chain of provenance is not guaranteed. You must add the visible disclosure required under Art. 50(4) yourself as the deployer.

Are the Content Credentials from DALL·E or Firefly enough?

They cover the machine readable level, as long as the file is not recompressed. However, Content Credentials do not satisfy the visible labeling for deepfakes and deceptively realistic images, because they are invisible. Add a visible icon as well.

Do I have to label locally generated Stable Diffusion images?

Yes, if the image qualifies as a deepfake, that is, looks deceptively real. Local setups usually embed no marking at all, so for content that requires labeling you add both the visible label and the machine readable metadata yourself.

Does the generator's labeling survive the upload?

Often not. Many platforms recompress images and, in doing so, remove metadata and sometimes C2PA as well. The visible icon burned into the pixel image, on the other hand, survives. Check after the upload which levels were preserved.

Which icon motif do I use for an AI generated image?

For an image fully generated by a prompt, use the "fully AI generated" motif. If only a part of a real image was generatively altered (inpainting, background), use the "partially AI modified" motif.

Is there a free tool that labels generator images?

Yes. The free web tool from AI labeling sets the official EU icon, embeds IPTC/XMP (optionally C2PA and a watermark), and keeps a proof register, regardless of which generator the image comes from.

Sources

• EU Commission, guidelines on transparency obligations: https://digital-strategy.ec.europa.eu/en/library/guidelines-transparency-obligations-providers-and-deployers-ai-systems
• EU Commission, EU icons for labeling AI generated content: https://digital-strategy.ec.europa.eu/en/policies/eu-icons-labelling-ai-generated-content
• Article 50 of the AI Act: https://artificialintelligenceact.eu/article/50/
• C2PA, Content Credentials: https://c2pa.org/

This article does not constitute legal advice. Please consult legal counsel for your specific situation.