Article 50 Is in Force: Who Labels What, and Who Pays
The transparency chapter applies since August 2, with no size threshold. The provider/deployer split, the editorial exemption, and a six-point readiness check.
Since 2026-08-02, the transparency chapter of the AI Act applies, with no size threshold and no exemption for small organizations. Non-compliance sits in the 15-million-euro or 3% penalty band, whichever is higher. For SMEs and start-ups, Article 99(6) caps the fine at whichever is lower. Here is the working split of who owes what.
If you provide an AI system that talks to people, it must tell them they are talking to a machine, unless that is already obvious from the context. Your customer-support bot needs to introduce itself as one.
If you provide a system that generates content, its outputs must carry machine-readable marking, metadata or watermark, identifying them as artificially generated. That marking duty sits with the tool provider.
If you publish, the visible duty is yours. Deepfakes and synthetic media you distribute must be visibly labeled. AI-generated text published to inform the public on matters of public interest must be labeled as machine-generated. The split is worth reading closely: the technical marking inside the file is the tool provider's job, while the visible label for the human who reads or watches belongs to the publisher, and it stays there even when the tooling comes from a vendor.
If you deploy emotion recognition or biometric categorization where it remains legal, you must inform the people exposed to it.
The Exemption Worth Designing For
There is one door out of the text-labeling duty, and it is instructive. AI-generated text that has passed genuine human editorial review, with a person holding editorial responsibility for it, escapes the labeling obligation. The law is pricing something here: human accountability is the compliance. A review step that exists only on paper is hard to evidence; a review step built into how content leaves the organization produces its own evidence.
Two more points of context. The Commission's enforcement powers over general-purpose AI models, including fines under Article 101, switched on the same day, 2026-08-02. And on marking formats, Europe chose a voluntary code of practice where China imposed binding labeling norms in September 2025, a divergence noted in the analysis this article builds on. The obligation is binding; the technical format is still being standardized. Plan for formats to tighten.
A Six-Point Readiness Check
- Every customer-facing conversational system introduces itself as a machine.
- Your content pipeline can show, for each published asset, whether it is synthetic and how it is labeled.
- The visible-label duty is assigned to a named internal team; vendor contracts do not carry it.
- Editorial review for public-interest text is a real gate with a named accountable person, so the exemption holds.
- Anyone exposed to emotion-recognition or biometric-categorization deployments, where legal, is informed.
- You keep evidence of all of the above, dated.
We spend our days building review gates of exactly this shape, so we may be biased toward them. In our BAMS control plane, an external communication can be configured to pass a human approval path that leaves evidence: who reviewed, what changed, when it shipped. Editorial responsibility implemented as plumbing produces dated evidence; implemented as policy alone, it produces an assertion.
On the data side, transparency work tends to expose a second problem: generation pipelines quietly carrying personal data. Runelm, our fail-closed sanitization proxy, classifies and pseudonymizes outbound data before it reaches a model, and blocks rather than sends cleartext when classification fails. Its community source is public on GitHub.
For completeness: BAMS enforces the review path rather than labeling content itself, Runelm is pre-1.0 and source-available rather than on PyPI, and this article is legal information, not legal advice.
Sources: Regulation (EU) 2024/1689, Arts 50, 99 (incl. 99(6) SME cap), 101; voluntary GPAI code of practice; China labeling norms (Sept 2025) as reported in "Le véritable problème de l'AI Act", Le SamourAI, 2026-08-03.