Aktiver — semantic AI platform Deterministic · traceable · behind your firewall
Insights / Regulation

EU AI Act transparency rules take effect while high-risk duties slip to 2027

On 2 August 2026 the transparency obligations in Article 50 of the EU AI Act became generally applicable and enforceable by national authorities across the Union. The date had long been described as the moment the Act’s full high-risk regime would arrive. It was not. Six days earlier, on 27 July, the Digital Omnibus on AI (Regulation (EU) 2026/1744) entered into force and moved the high-risk deadlines back by well over a year.

The result is a split calendar. Disclosure and labelling duties apply now, to almost anyone shipping generative or conversational AI into the EU market. The heavier obligations on risk management, data governance, documentation and human oversight for high-risk systems now start on 2 December 2027 for standalone Annex III systems and 2 August 2028 for AI embedded in regulated products.

/ 01What took effect, and what moved

According to Goodwin’s analysis, Article 50 now requires three things regardless of a system’s risk class:

  • Chatbots and virtual assistants must tell people they are talking to an AI system, unless that is obvious from context.
  • Synthetic audio, image, video and text, including deepfakes, must be marked by providers in a machine-readable, detectable format, and deployers must disclose deepfakes that depict real people, places or events.
  • Emotion recognition and biometric categorisation deployers must inform the people exposed to those systems.

There is one grace period: generative systems already on the market before 2 August have until 2 December 2026 to meet the marking obligation. Non-compliance with the transparency rules can draw fines of up to €15 million or 3% of worldwide annual turnover, whichever is higher. The Commission adopted guidelines on Article 50 on 20 July; TechTarget reports they run to 51 pages and make clear that content created before 2 August does not need retroactive labelling.

The same date also gave the Commission enforcement powers over the general-purpose AI rules that have applied since August 2025, covering technical documentation, copyright policies and systemic-risk assessment. Meanwhile the Omnibus moved Annex III high-risk systems, such as many hiring and credit-scoring uses, to December 2027, and product-embedded systems such as medical devices and machinery to August 2028.

/ 02Why the delay is not a reprieve

It is tempting to read the Omnibus as breathing room. The lawyers and governance specialists quoted by TechTarget mostly read it the other way. Andreas Splittgerber of Reed Smith said that if the high-risk requirements had started on 2 August, “almost any organization out there wouldn’t have been prepared.” Andrew Gamino-Cheong of Trustible pointed out that “the standards needed to demonstrate compliance with the Act’s high-risk requirements are still being developed.” The substance of the obligations did not change; only the start dates did.

What has changed is that the first enforceable layer is now live, and it is the layer that touches the most systems. A company may have a handful of high-risk use cases. It may have dozens of chat interfaces, summarisers and content generators, each of which now needs a disclosure or a machine-readable mark. Splittgerber’s advice was to start with an inventory: “A good inventory is really the first step, partly so that you can know what’s in scope or not.”

/ 03The inventory problem is a data-model problem

An AI inventory sounds like a spreadsheet. In practice it is a set of relationships: which model sits behind which product, which deployer exposes it to which users, which outputs are synthetic, which systems perform emotion recognition, and which of them will fall into Annex III in 2027. Each of those facts has a different owner, and they change every time a vendor updates a model or a team adds a feature.

That is the kind of information a knowledge graph handles well and a spreadsheet handles badly. When systems, models, vendors, use cases and legal categories are typed entities with explicit links between them, the question “which of our deployments generate synthetic images for EU users and have not yet implemented marking?” becomes a query rather than a survey. When the Annex III date arrives, the same graph can be extended with the high-risk classifications instead of rebuilt.

Traceability matters for the second half of the job too. Article 50 asks for outputs to be identifiable as AI-generated. Later obligations will ask for logs, documentation and evidence of human oversight. Systems that already record where an answer came from, and which data and rules produced it, will have much less to retrofit than systems that treat the model as a black box and the output as the end of the story.

/ 04What to watch and what to do

  • 2 December 2026. The marking grace period ends for generative systems already on the market, and new prohibitions on so-called nudification apps and systems capable of generating child sexual abuse material take effect, according to Goodwin.
  • Harmonised standards. Watch the European standards work on high-risk requirements; until it lands, conformity will rest on interpretation.
  • National authorities. Article 50 is enforced country by country. Early enforcement priorities will show how strictly “obvious from context” is read.
  • Your own inventory. Map every system that talks to people or produces content, record the model and vendor behind it, and record which obligations apply. Keep that map current rather than annual.

For teams already working with knowledge graphs, the practical step is modest: model the AI estate itself as part of the graph, so that compliance questions are answered from the same governed, traceable source as everything else.

How it unfolded
AUG 2025General-purposeAI rules apply[1]JUL 20, 2026Article 50guidelinesadopted[2]JUL 27, 2026DigitalOmnibus entersinto force[1]AUG 2, 2026Transparencyobligationsbecome enforc…[1]AUG 4, 2026This postDEC 2, 2026Marking graceperiod ends[1]UPCOMINGDEC 2, 2027Annex IIIhigh-riskrules apply[1]UPCOMINGAUG 2025General-purpose AI rules apply [1]JUL 20, 2026Article 50 guidelines adopted [2]JUL 27, 2026Digital Omnibus enters into force [1]AUG 2, 2026Transparency obligations become enforceable [1]AUG 4, 2026This postDEC 2, 2026 · UPCOMINGMarking grace period ends [1]DEC 2, 2027 · UPCOMINGAnnex III high-risk rules apply [1]
Aktiver’s take — Financial services

The Omnibus moved the hard deadlines, not the hard work. A bank that cannot list which of its systems talk to customers or generate content this year will struggle to evidence controls on its credit-scoring models in December 2027.

Treat the AI estate as data. In a graph, each customer chatbot, letter generator and scoring model is an entity linked to the vendor model behind it, the channels through which it reaches EU users, the Article 50 duty it triggers and, later, its Annex III classification. “Which deployments produce synthetic text for EU customers without a disclosure or machine-readable mark?” becomes a query a compliance officer reruns every week, with the same answer for the same state of the estate. When the high-risk rules arrive, those entities gain risk-management, logging and oversight links instead of a new spreadsheet. Because each link points back to the contract, design document or test record that asserted it, the inventory doubles as the audit file.

How we’d model it

AI systemFoundation modelArticle 50 obligationCredit scoring modelAnnex III high-riskuseSynthetic outputMachine-readable labelpoweredBysubjectToclassifiedAsmarkedWithAI systemFoundationmodelArticle 50obligationCreditscoring mod…Annex IIIhigh-risk u…SyntheticoutputMachine-readablelabelpoweredBysubjectToclassifiedAsmarkedWith
  • AI system —poweredBy→ Foundation model
  • AI system —subjectTo→ Article 50 obligation
  • Credit scoring model —classifiedAs→ Annex III high-risk use
  • Synthetic output —markedWith→ Machine-readable label

See where this applies →

Sources — read the originals
  1. Not Delayed, Not Deferred: EU AI Act Transparency Obligations Are Now in Force — Goodwin goodwinlaw.com ↗
  2. EU AI Act compliance deadline is here: What to watch — TechTarget techtarget.com ↗
Go deeper

See how Aktiver turns raw data into knowledge graphs your AI can reason over — deterministic, traceable, behind your firewall.

← All insights Subscribe via RSS → Book a demo →
Keep reading