AI-generated media

Every picture, video clip and synthetic voice that AI Tokens makes carries a mark inside the file that says it was generated by artificial intelligence. This page says which mark goes in which format, how to read it, what removes it, and why it is not enough on its own when you publish.

Why#

Article 50(2) of the EU AI Act, which applies from 2 August 2026, asks the providers of AI systems that generate synthetic audio, image, video or text content to mark it in a machine-readable form, so that it can be detected as generated. Text is covered too, not only media. The marks must be robust and interoperable, as far as the technology allows: the Code of Practice of June 2026 expects at least two layers, for example metadata plus a watermark.

Today AI Tokens adds one layer: metadata inside the file, described below, which is often lost when another program saves the file again (what removes it). There is no watermark, and chat answers and other text are not marked.

This page describes what we add to the files. It is not a legal assessment of your own obligations, which depend on what you publish and where.

The mark, format by format#

Each mark contains the same sentence:

Generated by artificial intelligence with AI Tokens (EU AI Act, art. 50).

The marks go on everything these services deliver, from the API and from the Studio in the dashboard alike: speech, and pictures and video where AI Tokens makes them. A picture or a clip refused by the content check is not delivered, so it is not marked either.

Text is not marked. A chat answer or a transcription is plain text, and nothing in it says how it was written.

How to read it#

A picture, in Python with Pillow:

from PIL import Image  # pip install pillow

picture = Image.open("picture.png")
print(picture.text["Comment"])
print("trainedAlgorithmicMedia" in picture.text["XML:com.adobe.xmp"])  # True

A clip or a voice, with ffprobe, which comes with FFmpeg:

ffprobe -v error -show_entries format_tags -of json clip.mp4
ffprobe -v error -show_entries format_tags -of json voice.mp3

The sentence appears as comment and description; for a WAV file as comment, with AI Tokens as encoder.

What keeps the mark, and what removes it#

The mark is metadata: it travels with the file as long as the file is not rewritten.

It stays when the file is copied, downloaded, attached to an email, put in a shared folder or sent as a file.

It can be lost whenever another program writes a new file:

• a screenshot or a screen recording is a new file, without the metadata;
• converting or compressing — a PNG saved as JPEG, a clip exported again by a video editor, an MP3 encoded again — keeps the mark only if the program copies these fields, and many do not;
• editing and saving in an image, video or audio editor: some programs keep XMP, few keep PNG text chunks or the comments of an audio file;
• social networks and messaging apps usually remove metadata when they publish a picture or a video; some read it first, to show a label.

There is no watermark in the pixels or in the sound, and no signed C2PA manifest. Whoever removes the metadata removes the mark, and nothing left in the file says where it came from.

A visible note is still needed#

The mark is for machines. A person who looks at a picture, watches a clip or hears a voice does not see it.

• Where you publish what you made — a web page, a social network, a presentation, a game — say it in words, next to it: for example Image generated with AI (AI Tokens). The Studio writes such a note under every result.
• If it could pass for real — a person, a place or an event that looks authentic — article 50(4) of the AI Act asks whoever makes it with the AI system, outside purely personal use, to disclose that it was generated or manipulated. For a work that is evidently artistic, satirical or fictional, a disclosure that does not spoil the work is enough.
• At school or at work, follow the rules of your course or of your organisation on the use of AI as well.