EU AI Act transparency rules hit on 2 August: what Article 50 actually requires

In under two weeks, telling people they are talking to a machine stops being good manners and becomes European law. Article 50 of the EU AI Act applies from 2 August 2026, and despite a wave of headlines about delayed deadlines, this one did not move.

Most of the EU AI Act's attention has gone to high-risk systems, and most of the recent news has been about deadlines slipping. Both facts have obscured a rule that lands on 2 August 2026 and applies to a very large number of ordinary businesses: Article 50, the transparency chapter.

What Article 50 requires

The obligations fall on providers and deployers of AI systems in specific situations that create transparency risk for people. In plain terms:

• Tell people they are dealing with an AI. Systems that interact directly with people must make that clear - unless it is already obvious in the relevant context. That qualifier is a genuine carve-out, not a loophole to lean on.
• Mark synthetic content machine-readably. AI-generated or manipulated audio, image, video and text must be marked in a form machines can detect.
• Disclose deepfakes. Deployers of systems producing deepfake content must disclose that it is artificially generated or manipulated, with exceptions for evidently artistic and creative work.

If you run a customer-facing chatbot in the EU, generate marketing assets with AI, or deploy a voice agent, this is your rule.

"Article 50 was not postponed. The high-risk deadlines moved; the transparency deadline did not."

Curious how AI engines describe your brand right now? Get a free visibility audit and see where you stand across ChatGPT, Gemini and Perplexity.

The postponement confusion, cleared up

You have probably seen headlines about EU AI Act obligations being delayed. They are real, and they are about something else.

The Digital Omnibus on AI - adopted by Parliament on 16 June 2026 and Council on 29 June 2026 - pushed high-risk obligations back: Annex III systems to 2 December 2027, and Annex I sectoral systems to 2 August 2028. Article 50 was not part of that.

The single change touching Article 50 is a grace period: systems already placed on the market before 2 August 2026 have until 2 December 2026 to comply with the machine-readable marking requirement in Article 50(2). Every other Article 50 obligation applies from 2 August. Do not let a colleague talk you into believing the date moved.

What it costs to get wrong

Article 99(4)(g) sets the penalty for breaching transparency obligations at up to 15 million euros or 3 percent of total worldwide annual turnover for the preceding financial year, whichever is higher. That is the mid-tier band - below the 7 percent reserved for prohibited practices under Article 5, above the 1 percent for supplying false information. For SMEs and startups, the cap is whichever figure is lower.

A practical checklist before 2 August

• Inventory your customer-facing AI. Chatbots, voice agents, automated email responders, in-product assistants. Anything that talks to a person is in scope.
• Check your disclosure is real. A machine-readable flag no one sees is not disclosure. The test is whether a reasonable person understands they are talking to an AI.
• Audit your generated media. AI-produced images and video in your marketing need marking. Work out who owns that step in your content pipeline.
• Do not over-rely on "obvious in context." It is a defensible exemption when a system is plainly a bot, and a weak one when your assistant is designed to feel human.
• Write down your reasoning. Where you conclude an obligation does not apply, record why. That record is what a regulator will ask for.

The takeaway

Article 50 is one of the least dramatic parts of the AI Act and one of the most broadly applicable. It does not ask you to prove your model is safe; it asks you to be honest about when a machine is involved. That is a low bar, which is precisely why missing it looks bad. Two weeks is enough time to inventory what you run and fix the disclosure gaps - and it is worth remembering that these rules govern how AI speaks to your customers, not how it speaks about you. The second question is still entirely yours to manage.

What this does not mean

Because Article 50 is short and broadly worded, it is easy to over-read. A few clarifications save a lot of wasted effort.

It does not ask you to publish your prompts, name your model, or explain how your system works. The obligation is disclosure of the fact that AI is involved, not disclosure of the mechanics behind it. A single honest line - "You are chatting with our AI assistant" - satisfies the interaction rule. You do not owe the user a technical briefing.

It does not turn every AI-touched asset into a labelled one. A marketing image that a designer generated and then heavily reworked by hand sits in a grey zone, and a product photo lightly retouched with AI tools is not the deepfake the rule is aimed at. The target is content that could pass for a genuine recording of real people or events. Ordinary editing is not the mischief here.

And it does not police what AI systems say about you. That distinction matters more than it looks. Article 50 governs the AI you operate; it is silent on the AI that answers questions about your company to everyone else. If a search assistant describes your product to a buyer, no transparency rule makes that description accurate or fair. That gap is yours to close, and it will not close itself.

A worked example

Picture a mid-sized software firm selling into the EU. It runs three things worth checking. First, a support chatbot on its help centre. Second, a monthly newsletter whose header images are generated with an AI tool. Third, a sales team that records short personalised video intros, some of which now use an AI voice clone of a founder who is short on time.

The chatbot is the easy case: add a clear opening line that the assistant is an AI, and the interaction obligation is met. The newsletter images need machine-readable marking at the point of generation, so the fix belongs in the content pipeline, not in the email tool. The founder voice clone is the one that bites. It produces synthetic audio of a real, identifiable person, which is squarely what the deepfake disclosure obligation covers. A visible note that the voice is AI-generated is not optional politeness here; it is the requirement.

"The obligation that catches people is rarely the chatbot. It is the thing that quietly imitates a real person."

The lesson from that firm is that the risky obligation is almost never the one you expected. The chatbot is obvious and gets handled early. The synthetic voice, the AI-generated spokesperson, the image that looks like a photograph - those slip through because nobody owns them. Assign an owner to each before 2 August, or the gap stays open by default.

How enforcement is likely to unfold

A rule taking effect is not the same as a rule being enforced from day one. Article 50 lands on 2 August 2026, but national market surveillance authorities need to stand up, staff up, and pick their early cases. The realistic near-term risk is less a surprise fine and more a complaint, a journalist's story, or a competitor pointing at an undisclosed bot. Reputational exposure tends to arrive before regulatory exposure.

That shapes how to prioritise. The obligations most likely to draw early attention are the visible, public-facing ones: consumer chatbots that hide their nature, and synthetic media of real people released without a label. Internal tools and low-stakes automation sit lower on anyone's list. Spend your two weeks where a member of the public could notice and object, because that is where the first cases will come from.

It also rewards keeping a paper trail. Where you decided an obligation did not apply, the written reasoning is what turns a tense enquiry into a short one. A firm that can show it thought the question through looks nothing like a firm that never asked it, even when they reach the same answer. Good records are the cheapest insurance on offer here.