The EU changed the rules for AI on 2 August. What you will actually notice
The AI Act's transparency rules became enforceable on 2 August 2026. What changed for people using AI, and the one deadline that has not arrived yet.
On 2 August 2026 the European Commission's AI Office and the national authorities working with it began enforcing the transparency rules in the EU AI Act. If you use an AI assistant in Europe, the honest report a month later is that your screen looks much the same as it did on 1 August. Nothing announced itself. No banner arrived to say your rights had expanded.
The rules did not fail. They were written for a different audience. Almost every obligation that took effect on 2 August is addressed to the company that builds or ships the AI system, and what a person sees is the residue of a duty owed by someone else. Working out which duty leaves a visible trace, and which leaves none at all, is most of what there is to understand here.
Three duties, all written for companies
The Commission's announcement, published on 31 July, names three changes that reach consumers. "Chatbots and other interactive AI systems will have to tell users they are dealing with AI, not a human." "Deepfakes (images, videos, or audio that have been edited or generated using AI) will have to be labelled." And "AI-generated or altered content will also have to carry machine-readable marks so it can be detected more easily." The stated purpose is to "reduce deception and manipulation and help people make informed choices."
Take them in that order and they descend from visible to invisible. The first you can read on a screen. The second you can read once whoever publishes the deepfake bothers to comply. The third you will never see at all, by design, because it is written for machines.
Why the disclosure rule changes so little
The chatbot duty sits in Article 50(1), which requires providers to ensure that 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". Article 50(5) adds that the information has to arrive "in a clear and distinguishable manner at the latest at the time of the first interaction or exposure". So far, so consumer-friendly.
Then comes the exception, and it is the reason your screen looks unchanged. The duty falls away where the situation is already "obvious from the point of view of a natural person who is reasonably well-informed, observant and circumspect". Open an app whose entire purpose is AI chat, choose a model by name from a list, and a regulator would struggle to argue you were under any illusion. Most chat interfaces in Europe were already inside the exception on the day the rule arrived.
Where it does bite is at the edges: a synthesised voice on a phone line, an assistant given a human name and a face, a support channel where you cannot tell whether a person has taken over. Those are the cases the drafters had in mind, and where you should expect wording to appear over the coming months.
The mark you cannot see has its own deadline
Article 50(2) is the ambitious one. It requires providers of systems generating synthetic audio, image, video or text to ensure the outputs "are marked in a machine-readable format and detectable as artificially generated or manipulated", with solutions that are "effective, interoperable, robust and reliable as far as this is technically feasible".
In practice a machine-readable mark is a signal buried in the thing itself. For an image it is usually a signed record in the file's metadata, or a pattern woven into the pixels that survives resizing. For text there is no metadata to hide in, so the marking works by nudging the model's choice between words that would have been equally good, in a pattern only someone holding the key can measure. You cannot see any of it, and that is the point: the mark exists so a detection tool or a regulator can ask a question about a file later.
It is also the one duty of the three that did not fully start on 2 August. The Commission's own guidance on Article 50 sets out a carve-out that most coverage skipped:
A limited grace period is envisaged only for AI systems placed on the market before 2 August 2026 and only as regards the marking and detection obligation for AI-generated content (Article 50(2) of the AI Act). Providers of such systems must comply with those obligations only as from 2 December 2026.
That covers almost every assistant you already use, because almost all of them were on the market well before August. MultiChats is one of them, and 2 December 2026 is our operative date for marking rather than 2 August. We have not shipped text marking, and saying otherwise four months early would be precisely the kind of overstatement these rules exist to discourage. If you read anywhere that every AI answer in Europe now carries a detectable mark, that claim is running ahead of the law.
Some providers moved early regardless. Anthropic shipped a text watermark for Claude on 14 August and named the AI Act as the reason, a mechanism worth its own explanation. The Commission also has a voluntary route running alongside the hard deadline: about 190 organisations had signed the Code of Practice on Transparency of AI-generated Content by the end of July, among them Anthropic, Google, Meta, Microsoft, Mistral and OpenAI. Signing replaces nothing, but the Commission frames it as "a streamlined, predictable and legally certain pathway to guarantee and demonstrate compliance", and the first task forces are due in September.
27 July pulled in two directions
Five days before the transparency rules landed, the AI Omnibus entered into force, described by the Commission as "a targeted simplification of the AI rulebook while preserving strong safeguards for people's safety and fundamental rights". It did two things that matter to a reader, and they point opposite ways.
It added a prohibition. The Omnibus "prohibits AI systems that generate non-consensual sexually explicit and intimate content or child sexual abuse material". Note the shape of that. The prohibition attaches to the system itself, so a company cannot discharge it by taking images down once complaints arrive. After a year in which image generators produced a steady supply of exactly this material, it is the most consequential consumer-safety line the EU wrote in 2026.
Then it delayed. The rules for high-risk systems in the sense the Act uses (hiring, credit, education, essential services, and the rest of Annex III) now apply from 2 December 2027, and high-risk AI embedded in physical products moves to 2 August 2028. The layer of the Act that governs decisions made about you, as opposed to disclosures made to you, is now well over a year further away than originally written.
Do not confuse this with the ChatGPT search designation
Two European instruments are running at once and almost every article blends them. On 31 August the Commission designated ChatGPT a Very Large Online Search Engine. That designation, and everything now being said about advertising repositories, belongs to the Digital Services Act, a separate regulation with its own clock and its own compliance deadline in January 2027.
The AI Act does none of that work. It places no obligation whatever on the sponsored card under a ChatGPT answer, and the Commission's enforcement announcement for 2 August does not mention advertising once. If a piece credits the AI Act with forcing an ad repository, it has merged two regulations drafted years apart.
Where that leaves us
We are based in Brussels, so this is home law for us, and one more thing belongs beside the December marking date. The transparency question people actually put to us is about training. What we can say and you can check: we do not train on your conversations. Some of the providers behind the models might, depending on which one you pick, and each model's information card says so in the picker before you send anything. Grok's is the bluntest: "xAI may use your conversations to train Grok models." Our privacy policy sets out the rest, provider by provider.
So is any of it worth anything?
Modestly, and mostly later. The disclosure duty was largely satisfied on arrival because the exception for the obvious swallowed the common case, and its real work will happen in voice and in human-seeming personas, which is where deception was always likeliest. The deepfake labelling duty asks the honest to label, and hostile actors will not, so it will help with commercial and casual synthetic media and do nothing against a Telegram channel pushing a fabricated politician.
The marking obligation is the one to watch, and the correct posture towards it today is patience. It is the hardest of the three technically, it is unfinished across the industry, and for products you already use it does not fully bite until 2 December. When it does, the change will still be invisible to you personally. What it buys is a world in which the question "was this generated" has a checkable answer somewhere, held by people who can act on it, rather than being settled by whoever sounds more confident in the replies. That is a smaller win than the headlines suggested in August. It is also the only one of the three that could not have been achieved by asking companies nicely.