The EU Wants You to Know When It Is AI

Some rules arrive with a fanfare. This one arrived on a Sunday in August, and six days earlier the law it belongs to was amended.

On 2 August 2026, Article 50 of the EU AI Act started to apply. Its concern is transparency. When people are talking to an AI system, or looking at content an AI system has made, the Act sets out who has to make that clear, and how.

This piece describes what the EU brought in and whom it falls on. It does not say whether any of it reaches a UK firm. That is a question of territorial scope under Article 2 of the Act, and it belongs with a lawyer rather than an article.

When it started, and the amendment six days before

The AI Act is Regulation (EU) 2024/1689. It applies in stages, and Article 113 sets the timetable. The general date is plain:

"It shall apply from 2 August 2026."

Article 50 sits in Chapter IV of the Act, and Chapter IV is not one of the parts Article 113 gives a different date. So 2 August is when it began.

On 24 July, the EU published Regulation (EU) 2026/1744, the Digital Omnibus on AI. It entered into force on 27 July. Among other things, it pushed the high-risk rules back to 2 December 2027 and 2 August 2028, depending on the kind of system.

The Omnibus affects Article 50 in just two ways. Paragraphs 1 to 6 were not changed at all. Read the amended text against the 2024 original and they match word for word. The Omnibus replaced paragraph 7, which deals with codes of practice, and it added a transition to a different article, Article 111:

"Providers of AI systems, including general-purpose AI systems, generating synthetic audio, image, video or text content, that have been placed on the market before 2 August 2026 shall take the necessary steps in order to comply with Article 50(2) by 2 December 2026."

So there is one delay. It covers one paragraph, and it helps one group: providers whose content-generating systems were already on the market. Every other duty in Article 50 started on 2 August. 2 December 2026 is now a date in this story, a year before the first of the new high-risk dates.

Two kinds of duty, two kinds of firm

Article 50 is easiest to read once you notice it is written for two different parties. The Act defines both in Article 3, and the definitions do the sorting.

A provider, under Article 3(3), is the person or body that develops an AI system, or has one developed, and places it on the market or puts it into service "under its own name or trademark, whether for payment or free of charge".

A deployer, under Article 3(4), is:

"a natural or legal person, public authority, agency or other body using an AI system under its authority except where the AI system is used in the course of a personal non-professional activity"

Put roughly, the provider builds or brands the system and the deployer uses it, other than in a personal capacity. Each gets its own paragraphs.

What falls on providers

The first duty is about conversation. Article 50(1):

"Providers shall ensure that AI systems intended to interact directly with natural persons are designed and developed in such a way that the natural persons concerned are informed that they are interacting with an AI system, unless this is obvious from the point of view of a natural person who is reasonably well-informed, observant and circumspect, taking into account the circumstances and the context of use."

Notice who the sentence is addressed to. The obligation is on how the system is designed and developed, so it lands on the provider. Notice too the exception. If it would be obvious to a reasonably well-informed, observant and circumspect person that they were dealing with AI, the duty does not bite. That is a judgement, not a checklist.

The second duty is about output. Article 50(2):

"Providers of AI systems, including general-purpose AI systems, generating synthetic audio, image, video or text content, shall ensure that the outputs of the AI system are marked in a machine-readable format and detectable as artificially generated or manipulated."

This is not a label a person reads. It is a mark a machine can detect. It does not apply where the system performs "an assistive function for standard editing", or does not substantially alter the input it was given. This is also the paragraph the Omnibus gave existing systems until 2 December to meet.

What falls on deployers

Deployers carry two duties of their own.

The first, in Article 50(3), covers emotion recognition and biometric categorisation systems. A deployer of either "shall inform the natural persons exposed thereto of the operation of the system".

The second, in Article 50(4), is where labelling comes in, and it has two halves. The first half is about deep fakes:

"Deployers of an AI system that generates or manipulates image, audio or video content constituting a deep fake, shall disclose that the content has been artificially generated or manipulated."

A deep fake, under Article 3(60), is content that "resembles existing persons, objects, places, entities or events and would falsely appear to a person to be authentic or truthful". Where the content is part of an evidently artistic, creative, satirical or fictional work, the duty narrows to disclosing that such content exists, in a way that does not hamper the display or enjoyment of the work.

The second half is about text, and it contains the most interesting sentence in the Article:

"Deployers of an AI system that generates or manipulates text which is published with the purpose of informing the public on matters of public interest shall disclose that the text has been artificially generated or manipulated. This obligation shall not apply where the use is authorised by law to detect, prevent, investigate or prosecute criminal offences or where the AI-generated content has undergone a process of human review or editorial control and where a natural or legal person holds editorial responsibility for the publication of the content."

Read the exemption slowly, because its two conditions are joined by "and". The text must have undergone human review or editorial control. And a person or organisation must hold editorial responsibility for publishing it. Review with nobody answerable for the result does not meet the wording.

It is not a ban, and it is not a label on everything. For text published to inform the public on matters of public interest, the Article offers two routes: tell the reader, or put the text through human review or editorial control with someone holding editorial responsibility for it.

Whichever duty applies, Article 50(5) says when the information is owed. It must be given "in a clear and distinguishable manner at the latest at the time of the first interaction or exposure".

The Code, and what it is worth

The Commission published a voluntary Code of Practice on marking and labelling AI-generated content on 10 June 2026, and guidelines on the Article 50 duties on 20 July.

It would be natural to treat the Code as the answer sheet. The Omnibus itself warns against reading too much into it. Its recital 41 says:

"Given that the codes of practice referred to in Article 50(7) and Article 56(6) have limited legal effect, and in particular do not grant a presumption of conformity, it is not strictly necessary for these codes to be approved by an implementing act."

Limited legal effect. No presumption of conformity. Following the Code is a route to meeting the duties, not proof that you have.

Meanwhile, in the UK

The UK has taken a different path, and it helps to start with a promise. In July 2024, the King's Speech said the government would "seek to establish the appropriate legislation to place requirements on those working to develop the most powerful artificial intelligence models".

The King's Speech 2026 followed on 13 May, with background briefing notes running to 129 pages and covering 37 bills. None of the 37 is an AI bill.

That is not the same as the UK doing nothing about AI. Some of what it has done sits in legislation already passed.

The Crime and Policing Act 2026 received Royal Assent on 29 April. Section 248 carries the heading "Power to amend Online Safety Act 2023: AI", and it came into force that day. It lets the Secretary of State amend the Online Safety Act 2023, by regulations, to minimise or mitigate the risks of harm to individuals in the UK presented by "illegal AI-generated content" and by "the use of AI services for the commission or facilitation of priority offences".

The operative word is illegal. The Act defines "illegal AI-generated content" as "illegal content that is AI-generated". This is a power aimed at content that is already unlawful, not a duty to tell people when something lawful was made by a machine.

Where AI does appear in the 2026 programme, it points another way again. The Regulating for Growth Bill would create sandbox powers, allowing existing rules to be temporarily relaxed, under strict controls, so that new products and technologies can be tested in real-world settings. Among its examples:

"Exploring cross-cutting AI sandboxes, enabling responsible testing and adoption of AI-enabled products and services across multiple sectors where existing regulatory frameworks currently slow innovation."

And the bill closest to home for anyone in financial services, the Enhancing Financial Services Bill, promises to make consumer protections "fit for the digital age". Its briefing note does not mention AI.

Put side by side, the two approaches are not really arguing about the same thing. The EU has written down, in advance, what people are owed when AI is used in ordinary, lawful activity. The UK measures set out here reach for the unlawful end of AI, and make room to test the rest.

Neither is obviously wrong. They are different bets about when you know enough to write the rule.

The question worth sitting with

The EU has chosen to require that certain AI announces itself. The King's Speech 2026 chose sandboxes, and no AI bill.

Will the UK stay on its own road, or will EU-style labelling arrive here anyway?

And would you wait to find out?