The machine now has to say what it is.
It still doesn’t have to know
who you are.
On August 2nd, the EU’s AI Act reached its main application date, and Article 50 became enforceable law: chatbots must disclose they are machines, synthetic media must carry machine-readable marks, deepfakes must be labeled — with penalties up to €15 million or 3% of worldwide turnover. Every disclosure the law requires runs in one direction: the machine identifying itself. Its recognition of you remains ungoverned.
On Sunday, August 2nd, Regulation (EU) 2024/1689, the AI Act, reached the date its own Article 113 sets for general application. Among the provisions that moved from text to enforceable law: Article 50, the transparency obligations. From this weekend, an AI system that interacts with people in the Union must be built so that they know it is an AI system. Synthetic audio, image, video, and text must be “marked in a machine-readable format and detectable as artificially generated or manipulated.” Deepfakes must be disclosed — and so must AI-generated text published to inform the public. The ceiling, under Article 99(4): fifteen million euros, or three percent of total worldwide annual turnover, whichever is higher.
Here is the statute, precisely. The chatbot duty applies unless the machine’s nature is “obvious from the point of view of a natural person who is reasonably well-informed, observant and circumspect, taking into account the circumstances and the context of use.” The marking duty itself carries no feasibility carve-out; what Article 50(2) qualifies is the technical solution behind it, which must be “effective, interoperable, robust and reliable as far as this is technically feasible” — and Article 50(7) obliges the AI Office only to “encourage and facilitate the drawing up of codes of practice” for detection and labeling, which means the law’s enforcement can arrive ahead of its technical standards. Artistic, satirical, and fictional work carries only a limited disclosure duty. And for SMEs and start-ups, Article 99(6) quietly inverts the penalty formula: fifteen million or three percent, whichever is lower. The law is narrower, more conditional, and more exact than its reputation — and the exact version is the one that belongs on a record.
One exemption deserves its own paragraph. The disclosure duty for AI-generated text published to inform the public does not apply where the content “has undergone a process of human review or editorial control and where a natural or legal person holds editorial responsibility for the publication of the content.” The European Union just encoded, in enforceable law, where accountability for machine-touched text lives: with the person who signs it — not the tool that drafted it. Editorial responsibility is now, formally, the difference between content that must announce its origins and content that stands on its editor’s name.
What the law does not reach
A compliant system can still misdescribe you. Article 50 governs the system’s nature (artificial, and disclosed) and its outputs’ provenance (marked, labeled). It says nothing about the factual fidelity of what those outputs claim about real people and firms.
Marking makes content traceable, not true. A machine-readable mark records where content came from. It does not record whether the content correctly identified the physician, the firm, or the thirty-year record it was describing.
The mirror is one-way, and now it’s law. The machine must be legible to you. Nothing yet requires you to be legible to the machine — and the systems composing answers about you read only what is machine-legible. The clinical name for that gap is Trust Transfer Failure.
A system can comply in full (introduce itself, mark its outputs, label its deepfakes) and still misidentify a specialist, misdescribe a firm, or assemble a recommendation set from which the qualified professional is structurally absent. The nearest regulatory movement on that side of the line is a proposed policy statement on AI output accuracy at the US Federal Trade Commission, whose comment window closed two days before Article 50 switched on — a docket this publication knows from the inside: the argument that accuracy includes recognition appears nowhere in the proposal, so Agentics filed it there as a public comment. Two continents, one summer: both turned to the machine’s honesty, one compelling it now, the other still deciding. Neither has reached the machine’s recognition.
The side of the line you can build on
Trust Transfer Failure, in full: the indicators of reliability that govern the physical world — board certification, licensure, standing, a verifiable record — never become machine-readable, and so carry no weight at the moment of recommendation. No parliament has scheduled that problem. It has no application date to wait for. It is an infrastructure decision, made or defaulted one entity at a time — and the build itself is no mystery: identity made machine-readable, signals consolidated where engines read, claims a system can verify. The Act took just over two years to travel from the Official Journal to enforceable law, all of it spent making the machines introduce themselves. Making the machines recognize you was never on the docket. That part was always yours to build.
Regulation (EU) 2024/1689 (the AI Act), Official Journal of the European Union — Article 50 (transparency obligations), Article 99(4) and 99(6) (penalties), Article 113 (application from August 2, 2026), quoted verbatim from EUR-Lex, CELEX:32024R1689.
European Commission, “Safer and more transparent AI,” August 2, 2026 — commission.europa.eu. This edition relies on the statute and the Commission’s announcement only; secondary-press characterizations of extended or delayed marking deadlines were checked against both and are not carried here.
Agentics LLC’s public comment on the US FTC’s proposed AI-accuracy policy statement, posted to the federal record July 31, 2026 — regulations.gov, FTC-2026-0859-0249.