The AI Act deadline that was not postponed

If you read anything about the EU AI Act in the spring, you probably read that it had been delayed. That was half the story.

In May, EU lawmakers provisionally agreed to push the high-risk obligations out, and the amending regulation that followed fixed the new dates at December 2027 and August 2028. A lot of firms filed the whole thing under "next year". But the transparency obligations in Article 50 kept their date. They applied from Sunday, 2 August 2026. The one narrow carve-out is for providers of generative AI systems placed on the market before 2 August 2026, who have until 2 December 2026 to comply with the machine-readable marking obligation. They are short, they are readable, and three of them touch the ordinary business of running a law firm's website.

None of this is dramatic. That is rather the point. Most of what Article 50 asks for is what a careful firm would want to be doing anyway. It is worth twenty minutes of your attention this week so you can stop thinking about it.

The three parts that reach a firm's website

One. The chat assistant on your site normally has to say it is an AI.

If a visitor is interacting directly with an AI system, they have to be told, unless it is obvious. The information has to be clear and easy to notice, and it has to arrive at the first interaction, not in a privacy policy three clicks away.

The legal obligation to build that in sits with whoever provides the assistant rather than with the firm using it. A firm that commissions an assistant as its own can end up wearing both hats. In practice that distinction is cold comfort, because it is your name above the conversation. Open your own site, start a chat, and see what it says in the first message. If it does not identify itself as an AI, that is a question for your supplier this week.

Two. AI-generated deepfakes have to be disclosed, and that one is on you.

If you publish an AI-generated or AI-manipulated image, audio or video that resembles a real or plausibly real person, place, object or event, and could falsely appear authentic to its intended audience, it may constitute a deepfake and must be disclosed. Clearly stylised or obviously illustrative content will not necessarily meet that test. That obligation sits on the deployer, which means the firm, not the tool. An obviously stylised illustration is one thing. A fabricated image passed off as your actual office, your people or a real case is another.

Three. AI-generated or AI-manipulated text, published to inform the public, has to be disclosed. Unless a human genuinely reviewed it.

This is the part worth reading twice.

Article 50 says that where a firm publishes AI-generated or AI-manipulated text for the purpose of informing the public on a matter of public interest, it has to disclose that. Then it gives an exemption, and the exemption is the whole story: the obligation does not apply where the content has gone through a process of human review or editorial control, and where a person or an organisation holds editorial responsibility for publishing it.

Two conditions, both required. The review has to be real. Commission guidance published on 20 July is explicit that a cursory look or a formatting pass does not count. It means someone competent read the thing, understood it, and had the authority to change it or stop it. And someone has to be answerable for publishing it.

Is a firm's practice-area guidance even in scope?

Honestly, it depends. The Commission's guidance names the administration of justice, fundamental rights and public services as matters of public interest, which puts some legal guidance in scope without settling every page.

The trigger turns on the purpose of publishing rather than the subject alone. A page about your opening hours is not informing the public on a matter of public interest. A plainly written explanation of what the two-year limitation period means for someone who was injured last spring is a much closer call, because that is exactly what it is doing.

Here is why the argument does not matter much in practice. If your content already goes through substantive review by a solicitor who can change it or refuse it, and your firm holds editorial responsibility for what appears on its own site, you are inside the exemption and the scope question never arises. If it does not, you are relying on winning an argument about the purpose of your own writing. That is a poor position for a regulated firm to choose voluntarily.

What a firm can do this week

1. Start a chat on your own website. Check that the assistant identifies itself as an AI in its first message, in plain words, not buried in terms.
2. Decide who holds editorial responsibility for the site, and write it down. It can be a named partner or the firm as an entity. What it cannot be is nobody.
3. Make review real, and make it recorded. If a solicitor reviews a page before it goes live, keep a record of who approved what and when. The Act does not prescribe an approval log, but a record is what lets you show the review and the responsibility existed in practice.
4. Review AI-generated imagery. Disclose it where it meets the deepfake test, or use photography and clearly illustrative content where the question does not arise.
5. Leave your archive alone. Content generated before 2 August 2026 does not need retroactive labelling, though the Commission encourages voluntary labelling where feasible.

On enforcement, Ireland has taken a distributed approach, with existing sectoral regulators designated as national competent authorities, and the AI Office of Ireland, established under the Regulation of Artificial Intelligence Act 2026, acting as the central coordinating authority and single point of contact. There is a penalty regime for Article 50 breaches with a ceiling of 15 million euro or 3 percent of worldwide annual turnover, though for small and medium firms the law caps fines at the lower of those figures, not the higher. For a professional services firm in Ireland, that number is not the realistic risk. The realistic risk is a client, a journalist, or a regulator asking who wrote something on your site, and the firm not having a clean answer.

The rule underneath the rule

Strip out the article numbers and Article 50 is asking one question. Who stands behind these words?

Irish solicitors have been answering that question for a very long time, first under the Law Society's rules and today under the LSRA advertising regime. The tools changed. What a name on a page means did not. That is genuinely good news for a firm that already treats its website as something it publishes rather than something a supplier posts to.

Two things follow, and they are the same thing. A page that has been read and approved by a solicitor who can refuse it is the version of your firm you would want a client to find. It also happens to be what the exemption in Article 50 describes.

This is not legal advice. We build the system, we do not advise on the law, and no one should act on a vendor's reading of a regulation. If Article 50 matters to how your firm publishes, take your own advice on it.

---

PageMax is the client-acquisition system built for law firms. Your named solicitors edit in plain words, approve and publish, with a record of who signed off on what. If you would like to see how that works on your own site, get early access. No obligation, and we will show you the workings.