# AI System Transparency Notice **Articles 50(1), 50(2) and 50(5), Regulation (EU) 2024/1689** This notice describes the AI systems the platform operator operates in the Intramedial service, and how we meet the transparency duties that apply to them. It concerns **our own** systems. It is not advice about yours. ## 1. You are interacting with an AI system **Number One** is an AI system. When you use it, you are interacting with a machine, not a person. It says so itself, before you exchange anything with it: the disclosure sits above the conversation, on every visit, in the clear and distinguishable manner Article 50(5) requires. It is not dismissible — a notice you cleared weeks ago informs nobody at the interaction it is meant to cover — and it is shown in your own language, carrying who operates the system and on whose instructions it acts. Number One can answer questions about your workspace, look records up deterministically, and — where you grant it permission — carry out actions in your workspace on your instruction. It is not a compliance determination, a legal conclusion or an audit sign-off, and it is designed to refuse rather than guess. ## 2. Machine-readable marking of generated text Where Number One's answer is composed by a language model, that output is marked in a machine-readable format and is detectable as artificially generated, as Article 50(2) requires. The marking is applied: - in the JSON response envelope and in HTTP response headers on the API; - as an in-band event on the streaming interface, delivered with the first token of generated text; - on the stored message, so a later read or export of the conversation carries the same marking rather than re-deriving it. The marker carries the IPTC digital source type `trainedAlgorithmicMedia` — the identifier used by C2PA and the wider content provenance ecosystem — so any consumer that already understands content credentials understands ours. **Where the marking does not reach.** If you select text in the interface with your mouse and paste it into another document, the marking does not travel with it. Carrying it would require embedding an invisible watermark in the prose itself. Any reformatting removes such a watermark, it corrupts text you may put in front of a regulator, and a hidden channel is the wrong thing to add to a product whose value is that nothing is hidden. Article 50(2) requires marking that is effective and robust *as far as technically feasible*; this is the edge of that, and we would rather state it than let you assume otherwise. **We do not over-mark.** Fixed text — a refusal, a quota notice, an error, a plan description — is written by people and selected by a program. It is not marked as artificially generated, because it is not. Every response carries an explicit origin so you can tell the difference rather than infer it. ## 3. Our own published material Where we publish text on this site that was generated or materially assisted by an AI system and concerns a matter of public interest, we disclose that on the page, as the second subparagraph of Article 50(4) requires. ## 4. Our role the platform operator declares itself both **provider** and **deployer** of the AI systems it operates in Intramedial, and discharges both duty sets. Our full role declaration, the per-system risk screening and the resulting obligations are published in the *EU AI Act Role Declaration* section of this page. ## 5. Contact legal@intramedial.com