Skip to content

Content marking

What the EU AI Act asks of an agent you build, what Fadenstack does for it, and what stays with you.

Who is responsible

Art. 50(2) asks the provider of an AI system that generates text, pictures, audio or video to mark its output in a machine-readable way, so that it can be detected as AI-generated. An organisation that builds an agent on the Fadenstack SDKs and uses it or offers it under its own name is, in our reading of the Commission's guidelines on Art. 50, the provider of that agent. The SDKs and the server's open-source core are components for building such systems: they are not subject to Art. 50 themselves, but they do the marking for you where they can.

Art. 50 applies from 2 August 2026. AI systems placed on the market before then have until 2 December 2026; anything released later has to be marked from the start.

What Fadenstack does

  • On the server, for every model call your agent makes: each complete text answer gets a signed receipt and a fingerprint, and answers of the organisation's own models also carry an invisible text watermark on a runtime that has it. Pictures get an invisible watermark, picture metadata and signed content credentials. The organisation's installation checks content for these marks. The administrators' side is in Help, under Content marking.
  • In the SDK, for what your tools place where people read it (a document, a mail, a web form, a file): the provenance module marks it in the way the host can carry (document properties or custom XML, mail headers, HTML attributes, file metadata) and has the server sign a record of each placement. See Content provenance for .NET and TypeScript.

What stays with you

  • Mark what your agent places. Declare what your tools output and place it through the provenance module. A host that can carry no mark is recorded as such, and the server's record is the mark then.
  • Keep the marks you find. When your agent edits a document or moves a picture, keep the metadata and content credentials it already has (Code of Practice, 1.2(a)). Insert a picture as the same bytes, and do not remove parts or tags that mark content.
  • Forbid removing them. Say in your terms of use that people must not remove, alter or hide the marks, and never ship or promote a tool that removes them (Code of Practice, 1.2(b)). For example: "Do not remove, alter or hide the marks that show content as generated by AI (watermarks, content credentials, metadata), and do not help others to."
  • Say that people deal with an AI (Art. 50(1)). An agent that talks to people tells them so, clearly and at the latest when they first meet it (Art. 50(5)). The SDKs' chat panels show such a notice. A mail your agent drafts for someone outside the organisation needs a visible line too: the machine-readable mark alone does not meet Art. 50(1).
  • Visible labels belong to whoever publishes (Art. 50(4)): a deep fake, or an AI-written text published to inform the public on matters of public interest, needs a label people can see, unless a person reviewed the text and someone holds editorial responsibility for it. You can offer your users an option to add one; the EU offers free icons for it.

What needs no mark

The Commission's guidelines leave out output nobody perceives as content: reasoning, actions such as a click or a web request, messages between agents, single words and labels, and source code with its comments. Edits that do not substantially change a person's text need no mark either: spelling and grammar fixes, light style polishing, translations, format conversion and transcription. A summary or a rewrite that changes style, structure and meaning does. Whether a change is substantial is decided case by case.

Not legal advice

This page describes what Fadenstack does and how we read the AI Act, the Commission's guidelines on Art. 50 and the Code of Practice on the transparency of AI-generated content. Whether your agent needs more depends on what it does and where.