What Article 50 requires
Article 50 sets transparency obligations. Its paragraph 2 covers providers of systems that generate images, audio, video or text: their outputs must be marked in a machine-readable format and detectable as artificially generated. Its paragraph 4 covers those who publish a deepfake: they must disclose that the content is artificial. The first part falls to Beemm Vision, the second to you. The official text is Regulation (EU) 2024/1689, Article 50.
3 formats
of images marked: PNG, JPEG and WebP, with no re-encoding
MP4, MOV
generated videos carry the same metadata
0 pixels
changed: it is metadata, not a watermark
Check the marking in four steps
Step 1
Download the original file
In the gallery, use the Download button: it saves the file exactly as Beemm Vision stored it, byte for byte. A screenshot, a copy-paste or an export from another app creates a new file, which may lose the metadata.
Step 2
Install a metadata reader
ExifTool is free and runs on Windows, macOS and Linux. Any tool that shows IPTC XMP metadata works too.
Step 3
Read the DigitalSourceType property
In a terminal, type exiftool -XMP-iptcExt:DigitalSourceType image.png. An image marked by Beemm Vision shows “Digital Source Type : http://cv.iptc.org/newscodes/digitalsourcetype/trainedAlgorithmicMedia”. Add -XMP-xmp:CreatorTool to read “Beemm Vision”. The same command reads an MP4 or MOV video.
Step 4
Make sense of an empty answer
The file was probably re-encoded (screenshot, export, social network), uploaded rather than generated, or generated before 7 November 2026, when marking arrived. Some model publishers also sign their files with a C2PA manifest: Beemm Vision leaves it intact, and a C2PA-compatible tool reads it.
What gets marked, at a glance
| File | Marking | Good to know |
|---|---|---|
| Generated PNG, JPEG or WebP image | IPTC XMP trainedAlgorithmicMedia | Inserted without re-encoding; a downsized variant is marked again |
| Generated MP4 or MOV video | Same XMP, at the end of the file | Playback does not change |
| File signed with C2PA by the model publisher | Left as is | Adding bytes would break its signature, which takes precedence |
| Other files: SVG, GIF, AVIF, audio, other videos | Not marked | Delivered as is: marking never blocks a delivery |
| Your uploads | Never marked | A file you upload is not a model output |
What you can do with it
Show where a visual comes from
Attach the exiftool output to a delivery: the metadata states plainly that the image comes from an AI model.
Edit without losing the label
Erase, upscale or remove the background in Beemm Vision: the result is a new generated image, marked in turn. Another app may strip it on export.
Label a post
On a social network, add “AI-generated image” or the platform’s own label: the metadata can vanish on re-encoding, and your audience does not see it.
Check a file you received
The same command reads any image: trainedAlgorithmicMedia flags a file declared as AI-generated. Its absence does not prove the file is authentic.
Frequently asked questions
- What does Article 50 of the AI Act require?
- Transparency. The provider of a system that generates images, audio, video or text marks its outputs in a machine-readable format, detectable as artificial (paragraph 2). Whoever publishes a deepfake discloses that the content is generated or manipulated (paragraph 4). The reference text is Regulation (EU) 2024/1689, Article 50.
- Is the marking a watermark?
- No. The AI Act marking Beemm Vision applies is metadata written into the file, without touching the pixels: the image looks exactly the same. Only a machine, or a tool such as exiftool, reads it.
- Does the marking survive an edit or a social network?
- Not always. Metadata disappears when a file is re-encoded without keeping it: a screenshot, an export from some apps, compression by some platforms. In Beemm Vision, a re-encoded variant is marked again; elsewhere, nothing guarantees it. Keep the original file.
- What is the Beemm Vision output filter?
- Before it is delivered, every generated image goes through a filter set up by Beemm Vision, independent of the model provider. An image flagged as explicit is delivered to no one. This filter blocks explicit content; it does not replace the marking. Videos do not go through it: they remain covered by the providers’ filters and the request checks.
- What is still up to me when I publish?
- Telling your audience that a piece of content is artificial when it could pass for real. If an image or a video looks like a real person, place or event, say that it is generated or manipulated; the Beemm Vision terms ask for this too (article 6). Never generate the image of a real person without their consent. The metadata does not replace this visible notice.
Generate a marked image
Open the app, generate an image, download it, then run exiftool -XMP-iptcExt:DigitalSourceType on the file.