Janet Xiu Shi | AI Quality Evaluation Advisor | Published 27 July 2026 (Updated 3 August 2026)
The short version: the delay headlines are about high-risk systems, not Article 50. Its transparency duties, including disclosing you are talking to a chatbot, have applied since 2 August 2026 as planned. Running a customer-facing chatbot or publishing AI-generated content? These duties are already in effect.
Article 50 of the EU AI Act sets transparency obligations for AI systems that interact with people or generate content. From 2 August 2026:
The Digital Omnibus on AI, Regulation (EU) 2026/1744, agreed by Parliament and Council in June 2026, published in the Official Journal on 24 July 2026 and in force from 27 July 2026, moved most high-risk system obligations to December 2027 and August 2028. That is real relief for providers of high-risk systems, and it is also where the confusion comes from: many teams read "AI Act delayed" and stopped preparing altogether.
The transparency duties in Article 50 were not part of that deferral. Neither was Article 4, the AI-literacy obligation, which is already in force and is covered in its own guide: AI literacy under Article 4, with inspection-ready evidence.
Ireland's central coordinating authority is Oifig IS na hÉireann, the AI Office of Ireland, established under the Regulation of Artificial Intelligence Act 2026, signed into law on 21 July 2026. It is an independent statutory body under the Department of Enterprise, Tourism and Employment, not a government department itself, and it does not enforce directly: it coordinates a distributed model of 15 designated National Competent Authorities, each enforcing within its own sector, including the Data Protection Commission, the Central Bank of Ireland, and the HPRA (designated under S.I. No. 366 of 2025 and DETE's 16 September 2025 designation notice). Section 70 of the Act lets an authorised officer require any person, by notice, to give information and to produce records, with no court order needed first. Paul Byrne was appointed the AI Office's first CEO on 30 July 2026, and the Department targeted the office to be operational from 2 August 2026. Once a question is asked, enforcement here runs on documentation: what gets requested is not whether you say you comply, it is whether you can show it.
The deadline is not the work - the architecture is the work. Use this three-step framework to move from deadline anxiety to actionable readiness.
The question: Where are you actually using AI today?
Not all AI carries the same obligations. The Act's real tiers are unacceptable-risk practices (banned outright), high-risk systems under Annex III (for example hiring and recruitment, and credit-scoring, subject to a narrower fraud-detection carve-out), the Article 50 transparency-only duties, and minimal-risk systems with no specific obligations. Article 50 is not itself a risk tier: certain AI systems that interact with people or generate content can pose specific risks of impersonation or deception "irrespective of whether they qualify as high-risk or not" (Recital 132), which is why transparency duties apply to them regardless of risk tier. Most customer-facing systems (chatbots, advisors, recommendation engines) sit in the transparency-only tier; decision-making systems such as hiring or credit-scoring tools may additionally trigger Annex III high-risk obligations, a separate and more demanding regime. Most organisations have not started this inventory yet, and mapping it is where readiness begins.
This is where the real work lives. Evidence is not a compliance checkbox. It is operational hygiene that makes your AI safer and more defensible.
You need:
Start with the highest-risk systems first. Document your processes. Test your controls. Then expand. Moving fast without audit trails is how you end up with liability; moving methodically with clear records is how you scale confidently.
Use these questions to surface the real operational issues (not just calendar anxiety):
I do not advise on Article 50 from theory. I built and operate Maya, an AI chatbot deployed as a teaching resource in an Irish state-funded AI programme, and I hold it to the same split I would apply to yours:
Alongside those sit 2 design choices that feed the same evidence trail an inspector asks for, but answer to the GDPR rather than Article 50: opt-in conversation logging, default off, deleted after 90 days, and chat content processed in AWS's EU infrastructure and stored in the EEA.
That is the level I review at. When I look at your chatbot I check the wording, the placement, the consent flow and the evidence trail, obligation by obligation and date by date, against the regulation and against a disclosure implementation I run in production myself. Two roles matter: the provider (who built the AI system and must mark generated content) and the deployer (you, who runs it and must disclose it to users). Article 50(1) is a deployer duty - it applies to you from 2 August, regardless of who built the system.
For teams running a customer-facing chatbot. You get a review of your Article 50(1) disclosure placement and wording, GDPR consent flow, and evidence trail, against the regulation and a working reference implementation. You end with a short written report: what is already aligned, what likely needs work, and the exact changes to make, including when the honest answer is that nothing needs to change.
Book a disclosure checkYes, from 2 August 2026. Article 50 of the EU AI Act requires that people interacting with an AI system such as a chatbot are informed they are interacting with AI, unless it is already obvious from context. The disclosure has to reach the user at the point of interaction, not sit in a policy page nobody opens.
The 2026 amendments deferred most high-risk system obligations to December 2027 and August 2028, but the Article 50 transparency duties were not deferred: chatbot disclosure, marking of AI-generated content, and deepfake labelling apply from 2 August 2026 (with a grace period to 2 December 2026 for the content-marking duty on systems already on the market). The Article 4 AI-literacy obligation is already in force.
Janet Xiu Shi, an independent AI Quality Evaluation Advisor based in Ireland, runs an Article 50(1) disclosure check in 5 working days, fixed fee: Article 50(1) disclosure placement and wording, GDPR consent flow, and the evidence trail against the regulation and against a reference implementation she operates in production.
Regulatory information current as of 3 August 2026. This page is practical guidance from an AI advisor: Janet Xiu Shi is not a solicitor and does not provide legal advice or legal services. For a legal opinion on your specific obligations, consult a qualified solicitor.