Article 50 of the EU AI Act applies from 2 August 2026 and most Irish organisations it touches still have nothing in place. The obligations are simpler than the commentary suggests: tell people when they are talking to AI, label AI-generated content that could mislead, and disclose emotion recognition and biometric categorisation. Here is what compliant looks like in practice — with the wording patterns that work.
Our earlier piece on Article 50's transparency obligations explained who is in scope and when. This one answers the question that follows in every conversation: what does compliant actually look like? With the 2 August 2026 application date days away, the gap between "we know about Article 50" and "we have implemented Article 50" is where most Irish organisations currently sit.
The chatbot disclosure
The obligation: people interacting with an AI system must be informed they are interacting with AI, unless it is obvious from context to a reasonably well-informed person.
What works in practice is a disclosure that is present at the start of the interaction, plainly worded, and not buried:
"You're chatting with our AI assistant. It can answer most questions about your account; you can ask for a human at any time."
That sentence does three jobs: it discloses, it sets capability expectations, and it offers an exit. The patterns that fail are the ones designed to technically-comply: AI branding hidden in a logo, disclosure in the terms and conditions, or a bot given a human name, a human photo, and typing indicators — a design that actively works against the disclosure it links to. If your interface is engineered to make users forget they are talking to AI, a footnote does not fix that.
The "obvious from context" carve-out is narrower than people hope. A tool named "AI Assistant" inside an enterprise platform is probably obvious. A conversational widget on your public website handling customer service is not — assume disclosure is required.
Labelling AI-generated content
Providers of generative systems must ensure outputs are marked machine-readable as AI-generated; deployers who publish AI-generated text on matters of public interest, or deepfake-class image, audio and video content, must disclose the generation. The four-month grace period the Digital Omnibus granted on machine-readable watermarking softens the provider-side timeline — it does not remove the deployer-side obligation to disclose.
Practical translation for the common Irish cases:
- AI-written blog or news-style content published without meaningful human editorial control needs a visible statement — a byline-level note such as "This article was generated with AI assistance" is the emerging convention.
- Substantially human-edited content where AI drafted and a person reviewed, revised and took responsibility falls outside the disclosure obligation for public-interest text. Document the editorial step; that documentation is your evidence.
- Synthetic media — a cloned voice on a phone line, an AI-generated presenter video — needs disclosure regardless of how benign the use is.
Emotion recognition and biometric categorisation
If you deploy systems that infer emotion or categorise people from biometric data, the people exposed to them must be informed. In the Irish market this shows up most often in HR and customer-analytics tooling bought as SaaS — sentiment scoring on support calls is the classic example. Two things to note: workplace emotion recognition is not just a transparency question, it sits close to the Article 5 prohibitions we covered in what still lands on 2 August; and "the vendor didn't mention that feature" is not a defence available to deployers.
Who enforces this against you
From 1 August, Ireland's enforcement architecture is live — the AI Office coordinating, with supervision distributed to sector regulators. We mapped which authority watches which use in Which Regulator Actually Supervises Your AI in Ireland, and the fuller picture of the domestic framework is on our AI Office of Ireland explainer. Article 50 breaches are the visible kind — anyone can screenshot an undisclosed chatbot — which makes them a plausible early-enforcement and complaint-driven category.
A ten-day implementation sequence
- Inventory the touchpoints — every chatbot, voicebot, generated-content channel, and emotion/biometric feature, including ones embedded in purchased tools.
- Draft the disclosures — one sentence per touchpoint, using the patterns above. This is hours of work, not weeks.
- Fix the dark patterns — human names and fake typing indicators on bots are now a liability, not a conversion tactic.
- Document the editorial line — decide and record which published content is AI-assisted vs AI-generated, and who signs off.
- Put it in the governance file — Article 50 evidence belongs in the same structure as your literacy and inventory work, which is exactly what our EU AI Act readiness review builds.
Transparency is the cheapest obligation in the entire Act to meet — and correspondingly hard to explain away when it is missing. If you want the touchpoint inventory and disclosure wording done with an independent eye, our EU AI Act consulting covers Article 50 as part of the broader readiness picture.