AI Act: what applies now and what it means for your organisation
Article 4 (literacy) has applied since February 2025. Article 50 (transparency) applies from August 2026, with a grace period to December 2026 for marking in systems already on the market. Romania had not adopted its national implementing act when ANCOM last spoke, but the obligations in the Regulation apply regardless.
The reference text is the consolidated AI Act of 27 July 2026, which includes the changes made by Regulation (EU) 2026/1744 (the Digital Omnibus on AI, Official Journal of 24 July 2026). Use it rather than the original 2024 timetable.
What applies now and what comes next
| Provision | Applicable from | What it requires |
|---|---|---|
| Art. 4: AI literacy | 2 February 2025 | Providers and deployers take measures to support the AI literacy of their staff and of other people who operate or use AI systems for them, taking account of knowledge, experience and context. Art. 4(1) does not require guaranteeing a specific level of literacy for any individual. The Commission FAQ says no certificate is needed, no governance structure is mandated and no training is imposed, and suggests keeping an internal record of training and guidance. |
| Art. 50(1): Interaction identification | 2 August 2026 | Providers design AI systems that interact directly with people so that the person is told they are dealing with an AI system, unless that is obvious to a reasonably well-informed, observant person. Background machine-to-machine processing is a different case. |
| Art. 50(2): Synthetic content marking | 2 August 2026 | Providers of AI systems that generate synthetic audio, image, video or text mark the outputs in a machine-readable format so they are detectable as artificially generated, as far as technically feasible. The obligation does not apply where the system performs an assistive function for standard editing or does not substantially alter the input or its meaning. Newly drafted text is generated, not lightly edited, so do not assume the exception covers it. |
| Art. 50(4): Public-interest text | 2 August 2026 | Deployers of an AI system that generates text published to inform the public on matters of public interest disclose that it is artificially generated. The exception requires that the text underwent human review or editorial control and that a person holds editorial responsibility for the publication. Internal policies are generally not published to inform the public. The text does not define the review needed, so do not treat an approval click alone as a safe basis. |
| Art. 111(4): Art. 50(2) grace period | Until 2 December 2026 | Providers of systems generating synthetic content that were placed on the market before 2 August 2026 take the necessary steps to comply with Art. 50(2) by 2 December 2026. It covers Art. 50(2) only. Whether a given system counts as placed on the market before that date depends on the facts. |
Provider or deployer? Why it matters for you
Under Art. 3(3) a provider develops an AI system, or has one developed, and places it on the market under its own name or trademark. Under Art. 3(4) a deployer uses an AI system under its authority. Art. 3(68) defines a downstream provider as a provider of an AI system that integrates an AI model, whether it is the provider’s own or comes from another entity under contract. A company that builds and sells its own AI drafting or triage feature on top of another company’s model is therefore likely a downstream provider. Customer organisations using that feature are deployers. Bring-your-own-key and self-hosting do not change this reading. Using a model does not by itself make a company a provider of that model.
Practical consequences:
- As provider: Art. 50(1) (interaction identification) and Art. 50(2) (marking of outputs).
- As deployer (your client): Art. 50(4) if the organisation publishes AI-generated text to inform the public on public-interest matters, plus Art. 4 (literacy).
- Do not conflate roles: your contract with a model provider is separate from your role as provider of an AI system.
Detectable marking: what is not enough
- A visible footer or warning label on a document is readable by people. It does not by itself show the machine-readable marking that Art. 50(2) names.
- Metadata alone is not shown to meet the obligation. Art. 50(7) provides for Commission-encouraged codes of practice, and possibly implementing acts, on detection and marking, so check them before choosing a format.
- Provenance (who generated, reviewed and approved, which version) and machine-readable origin are different things. Keeping both through revisions and exports is good practice.
High risk classification: not automatic
Drafting governance and compliance text and advisory security triage are not automatically high risk just because customers are NIS2 entities. Record the intended purpose and reassess if you move into areas such as employment decisions, biometric identification or safety functions in critical infrastructure.
“Nobody can be fined until Romanian law”: do not rely on it
ANCOM’s statement of 24 July 2026 says the AI Act applies from 2 August 2026 with exceptions, that the national act is still being drafted by the designated authorities, and that ANCOM and the other authorities can check and sanction non-compliance only after that act enters into force. ANCOM adds that its current role is to take part in preparing the framework and to inform the public, and it advises everyone who develops, places on the market or uses AI systems to review their obligations now. A parliamentary question of 14 September 2026 put to the Ministry of Economy, Digitalisation, Entrepreneurship and Tourism raises the missing national framework and the resulting inability to sanction. It is a question, not a legal ruling. We found no national act adopted after these two documents.
This is not immunity. The obligations in the Regulation apply from their dates whatever the state of national enforcement. The Commission states that national competent authorities enforce the rules for most AI systems, while the AI Office enforces them for general-purpose AI models and for systems built on a model from the same provider.
AI Act sanctions: ceilings, not automatic fines
| Violation | Ceiling |
|---|---|
| Prohibited practices (Art. 5), Art. 99(3) | EUR 35m or 7% of worldwide turnover, whichever is higher |
| Supplying incorrect, incomplete or misleading information to authorities, Art. 99(5) | EUR 7.5m or 1% of worldwide turnover |
| Transparency obligations (Art. 50), Art. 99(4)(g) | EUR 15m or 3% of worldwide turnover, whichever is higher |
| Art. 4 (literacy) | Not listed in Art. 99(4). Member States set the penalties (Art. 99(1)); the 15m/3% figure is not the automatic penalty for Art. 4 |
SMEs and small mid-caps: for SMEs, including start-ups, each fine is up to the percentage or the amount, whichever is lower (Art. 99(6)). Art. 99(6a), added by Regulation 2026/1744, applies the same lower-of rule to small mid-caps (SMCs) for fines under Art. 99(4) and (5). Art. 99(1) also asks Member States to take the interests and economic viability of SMEs and SMCs into account.
What this means for you, now
- Art. 4 is already live: take measures to support the AI literacy of the people who use AI for you. Keep a simple record of training, guides and briefings.
- Art. 50(1) and (2) apply from August 2026: if your system was on the market before 2 August 2026, the grace period to 2 December 2026 covers marking under Art. 50(2) only. Interaction identification applies from August.
- Work out your role: provider of an AI system (likely, if you build one on a third-party model) or deployer. A contract with a model provider does not remove your own obligations.
- Plan marking beyond a visible footer: Art. 50(2) names a machine-readable format. Follow the codes of practice and guidance the Commission issues.
- Do not assume national immunity: the missing Romanian act affects who can sanction and how, not whether the obligations apply.
- Consider keeping a provenance trail (generation, review, approval, export), and name who holds editorial responsibility if you publish AI-assisted text on public-interest matters.
What Phalanx provides around AI
Phalanx provides records and controls around the AI you use. It does not claim to make you compliant with Art. 50, and it does not mark AI output or track its provenance. Today it does the following:
- Each organisation can use its own AI provider key.
- Records flagged as personal data are never sent to an AI provider.
- AI-generated text in the policy editor carries a warning label asking the user to review the changes before saving.
- The platform refuses AI calls that the organisation’s licence state, module access or role scope does not permit.
- Every AI call is written to the audit log, and the entry is committed before the request is sent.
- An AI usage log records how AI features are used.
These records support your own literacy and oversight documentation. Deciding whether Art. 50 applies to your use of AI, and how to mark outputs, remains your assessment.