Engineering for the EU AI Act: Article 50 transparency as system architecture

On 2 August 2026 the transparency rules in Article 50 of the EU AI Act became enforceable, together with the Act's full penalty regime. Weeks earlier, the Digital Omnibus on AI, Regulation (EU) 2026/1744, rewrote the compliance timeline, entering into force on 27 July 2026, three days after publication, precisely so it would land before that date. Most published guidance, and most model training data, still describes the pre-amendment law. Building against it produces the wrong system.

This paper sets out how we read the amended Act as engineers, and the architecture we use to satisfy it across our products. Nure runs governed agents that interact with people directly on a business's behalf. Logi corresponds about quotes and exceptions in supply operations. Both reach people in the EU, so both carry Article 50 duties. We are a UK provider, and that changes nothing: the Act applies wherever a system's output is used in the EU. Our position is simple: a transparency duty is a property of the system, not a policy document. It is engineered in, or it does not exist.

1. What changed on 2 August 2026

The Digital Omnibus (Regulation (EU) 2026/1744, in force 27 July 2026) amended Regulation (EU) 2024/1689 and moved the dates that matter to engineering teams. The amended timeline:

The trap is stale sources. Well after the amendment, widely used references, including consolidated legal texts and implementation trackers, still served pre-amendment rules stating that Annex III high-risk obligations applied from August 2026. They do not; the Omnibus moved them to December 2027. Any source carrying the old date predates the amendment and should not drive a build decision.

Article 50 is enforced mainly by national market surveillance authorities, not the AI Office, and infringements carry fines up to €15m or 3% of worldwide turnover. The Omnibus also extended SME relief: simplified technical documentation, proportionate quality-management obligations and mitigated penalties now reach all SMEs, including startups.

2. What does not apply: most AI-native products are not high-risk

Annex III lists the stand-alone high-risk domains: biometrics, critical infrastructure, education, employment, essential private and public services (including creditworthiness), law enforcement, migration and the administration of justice. Freight quoting is not on the list. Agent-run business communication is not on the list. Content generation is not on the list.

That classification is load-bearing, because it defines what you should refuse to build. A product that is not high-risk does not need the Articles 9 to 17 apparatus: no risk-management system, no CE marking, no notified-body conformity assessment, no EU database registration. A backlog that accumulates those artefacts "to be safe" is spending its compliance budget on the wrong law. The duties that do attach are the transparency duties of Article 50, and they attach now, not in 2027.

3. Article 50(1): when the disclosure duty attaches

Article 50(1) requires that a person interacting with an AI system is told so. It is a provider duty, satisfied at design time, before the system is placed on the market. The duty attaches when four criteria are met cumulatively: the software qualifies as an AI system; it is designed for a genuine two-way exchange rather than one-shot output; the interaction is direct, not routed through a human intermediary; and the counterparty is a natural person. Classified against real product surfaces:

Two edges deserve respect. First, the human-intermediary exception requires substantive review. The Commission is explicit that formal or procedural checks, such as spell-checking and grammar, do not count. A UI that lets an operator approve twelve drafts in ninety seconds is not an exemption; it is a defect dressed as one, and the system should flag that pattern rather than accept it. Second, the "obvious to an average person" exception is to be interpreted restrictively. We do not rely on it.

Delivery is as constrained as applicability: the notification must arrive from the start of the first interaction, clearly and distinguishably, and meet accessibility requirements. A footer on message four fails. Disclosure copy that hedges ("may have been generated with AI assistance") fails differently: it discloses nothing. The recipient is interacting with an AI system, and the sentence says so.

4. Article 50(2): marking synthetic content

Providers of systems that generate synthetic audio, image, video or text must ensure outputs are marked machine-readably and detectable as artificially generated. Marks must be effective, reliable, robust and interoperable. The duty carries confirmed exclusions: short sequences of numbers or letters, source code, outputs exchanged machine-to-machine without human exposure, closed-loop industrial use that never becomes final output, and AI performing an assistive function for standard editing. The Commission's Guidelines on Transparency of AI-Generated Content (July 2026) add a narrow exemption for certain business-to-business and industrial contexts, subject to conditions set out in the guidelines themselves. That is a determination to make against those conditions per surface, never an assumption.

Two engineering consequences. The mark is applied at generation, not at export: it travels with the artefact from the moment the artefact exists, and it must survive the render and encode stages where marks typically get stripped, which is exactly where we test it. And the grace period is narrower than commonly read: only systems placed on the market before 2 August 2026 get until 2 December 2026, and only for the 50(2) marking duty. Nothing generated before 2 August 2026 needs retroactive labelling.

5. Article 50(4): deepfakes and public-interest text

Article 50(4) is a deployer duty with two limbs. The deepfake limb covers generated or manipulated image, audio or video that resembles existing persons, objects, places or events and would falsely appear authentic: three cumulative criteria. Disclosure must reach the person at first exposure at the latest, perceivable without technical tools, which means a visible or audible label. A provider's machine-readable 50(2) mark does not satisfy it. Whether a given output is a deepfake is content-dependent, not system-dependent, which argues for a classification step in the pipeline rather than a blanket rule.

The text limb covers only text that is published, informative to the public, and on a matter of public interest (politics, justice, public health and similar). It is exempt where the text underwent genuine human editorial control. Commercial correspondence and freight quotes fail those limbs outright: no 50(4) labelling duty attaches to them. Reading the limbs precisely prevents both under-building and the quieter failure of over-labelling everything until labels mean nothing.

6. Transparency as architecture: the invariants

We encode the duties as invariants: properties that hold across every code path, and that a build fails for violating even when the feature works. The load-bearing eight:

1. Disclosure is a property of the system, not of configuration. No tenant setting, template edit or environment variable can disable it. A code path that can emit a customer-facing message without disclosure is a defect.
2. First contact carries the disclosure. Asserted at the point of composition, not appended downstream.
3. Human-in-the-loop status is recorded, not assumed. Every outbound message records whether a person substantively reviewed it, who, and when. No record means no review.
4. Provenance marking happens at generation, not at export.
5. Model provenance is queryable. For any artefact: which model, which vendor, which version, at what time.
6. An unresolved legal question is a wall, not a default. Work that depends on it stays blocked and says so loudly; nobody infers around it.
7. Every compliance behaviour has a golden fixture. That includes a negative test proving the control cannot be bypassed. A behaviour without a failing-then-passing test is not implemented.
8. Jurisdiction is data, not a constant. Whether a recipient is in the EU is resolved per message and logged, never hardcoded per deployment.

The flow above is these invariants drawn as a pipeline: marking and registry at generation, the routing decision, disclosure or recorded review, and per-message jurisdiction at delivery. There is no arrow around the controls, because there is no such code path. The approval-lane discipline underneath, which agent actions need a human and how review is made cheap enough to be real, is the subject of Governed autonomy.

7. Sequencing the build

The order of work matters as much as the invariants. We start with an audit-only unit: enumerate every code path that produces output reaching a natural person outside the organisation, and record for each whether it is two-way or one-shot, direct or intermediated, its modality, and the recipient's jurisdiction. No other unit starts until that inventory exists and a human has reviewed it. Every subsequent decision keys off it, and an inventory inferred later is one that missed a path.

Then strictly one unit at a time, each in four phases: define the goal and its constraining invariants; attack the design before writing code, listing the ways it can be satisfied in form but violated in substance; audit what the code actually does today, reported plainly rather than quietly fixed; and only then build the smallest change, failing test first. Client-facing documentation (system description, model provenance, human-in-the-loop map, change log) is generated from the inventory and the registry, so it cannot drift from reality. Under Article 4, both providers and deployers must take measures to support AI literacy; the businesses deploying our products discharge their own duty partly with that material, so we treat it as a deliverable of the build, not an afterthought.

8. The national layer: Italy

The AI Act is a floor, not the whole law. Italy's Law No. 132/2025, in force since 10 October 2025, made it the first member state with a dedicated national AI framework, layered on top of the Act. AgID acts as notifying authority; ACN is the market surveillance authority and Italy's single point of contact with the EU. Two implementing decrees received preliminary approval in June 2026, covering authority powers, AI in education, policing biometrics and a new criminal-liability provision, but they are not yet in force. We do not build against pending texts; we track them.

The Italian layer matters for a structural reason: it adds civil and criminal exposure above the Act's administrative fines. That raises the defensive value of documentation and provenance records independent of whether the Act itself mandates them: the queryable registry earns its keep twice.

9. Sources

The authoritative references, in the order an engineering team should trust them. None of this paper is legal advice; dates and article references should be confirmed against the Official Journal text with qualified counsel before commercial reliance.

1. Regulation (EU) 2026/1744, the Digital Omnibus on AI: full text on EUR-Lex.
2. European Commission, FAQ on the Article 50 transparency obligations.
3. European Commission, Guidelines on Transparency of AI-Generated Content, July 2026.
4. European Commission, Code of Practice on Transparency of AI-Generated Content.
5. European Commission, enforcement framework of the AI Act.
6. OECD.AI, Italy, Law 132/2025 overview.

This paper is part of our Engineering research. Related work runs through the governance, agentic, system design and operations threads, and the full library lives at Research.