EU AI Act Article 50: Chatbots Must Now Say They’re AI

Four years after it was first proposed, the EU AI Act just crossed a threshold that actually matters for everyday life. As of 2 August 2026, any AI system that holds a conversation with you — a chatbot, a voice assistant, an AI agent on a website — must tell you it’s AI. Any deepfake must carry a visible label. Any AI-generated image or video must be marked in a way that detection tools can find. These are not proposals or pilot schemes. They are live legal requirements, and the fines for ignoring them are real.

UK investors keep asking me whether the EU AI Act matters post-Brexit. It absolutely does. If your business sells to EU customers, operates any service accessible in the EU, or deploys AI tools that interact with EU users, Article 50 applies to you. The borders on the map changed. The rules on your laptop screen did not.

What Is Article 50 and Why Does It Matter Now?

Article 50 is the EU AI Act’s transparency chapter. It covers AI systems that interact with people — not high-risk systems like medical devices or credit scoring tools, which face a different and more demanding set of obligations — but the everyday AI that millions of people encounter without realising: customer service bots, AI writing assistants, voice responders, and deepfake content generators.

The August 2 deadline was not delayed. A separate set of obligations under the Act — particularly those targeting high-risk AI — were deferred by the EU’s Digital Omnibus package earlier this year. Article 50 was explicitly carved out of those delays. It landed on schedule. That distinction matters because many compliance teams were tracking the broader postponements and may have assumed the whole Act was pushed back. It wasn’t.

When I looked into how the different provisions interact, the clearest way to think about it is this: Article 50 targets transparency, not risk. Its job is to make sure people know when they’re dealing with AI. That’s a lower bar than high-risk compliance, which is why regulators kept the August date.

The Chatbot Disclosure Rule

The most immediately practical obligation under Article 50 is for AI systems that interact directly with people. Providers must design systems so users know they’re talking to an AI — not a human — unless the context makes it completely obvious. The legal standard is disclosure “in a clear and distinguishable manner.”

This catches an enormous range of products. Customer service chatbots on retail and banking websites. AI voice assistants embedded in phone systems. Automated social media accounts. AI agents that respond to emails on behalf of businesses. Any system that holds a real-time conversation with a person needs to flag that it’s artificial.

There’s an exception for creative or entertainment contexts where a fictional AI persona is part of the product — but that exception is narrow. If a reasonable user might not know they’re talking to AI, the disclosure is required. “It should have been obvious” is not a defence that will hold up with regulators.

The Deepfake Labelling Rule

Article 50(4) covers deepfakes specifically. Deployers — the businesses and individuals who actually use AI tools to generate synthetic content, rather than the companies that build those tools — must label deepfakes when they share them. The label must appear at the point of first exposure, clearly and visibly.

The practical trigger is generating or manipulating “image, audio or video content constituting a deep fake.” That language is deliberately broad. An AI-generated video of a politician saying something they never said is clearly a deepfake. But so is a corporate promotional video that swaps one face for another, or an AI-voice clone used in an advertisement without disclosure. The Act doesn’t distinguish between malicious fakes and commercial ones. Both require labelling.

There is an exception for content used in art, satire, or fiction — but even then, if the content is likely to cause confusion about its artificial nature, the label still applies. The creative exemption is not a blanket pass for any content that calls itself artistic.

The Machine-Readable Marking Rule

Article 50(2) requires providers of generative AI systems — the companies building tools that generate images, video, text, or audio — to embed machine-readable markers in their outputs. These are technical watermarks detectable by other software, designed to let platforms and regulators identify AI-generated content even when it doesn’t carry a visible human-readable label.

This is where the grace period applies — and it’s important to understand exactly what the grace period covers. AI systems already placed on the market before 2 August 2026 have until 2 December 2026 to comply with the machine-readable marking requirement. That’s a four-month window for existing products to retrofit watermarking into their outputs.

What the grace period does NOT cover is the other Article 50 obligations. Chatbot disclosure, deepfake labelling, and emotion recognition notices apply from 2 August regardless of when the system was first deployed. An AI chatbot that went live in 2023 still has to disclose itself as AI right now. The December deadline only applies to the technical watermarking specification in Article 50(2).

What About Emotion Recognition and Biometric Systems?

Article 50(3) covers a less widely discussed category: AI systems that infer emotions or categorise people using biometric data. These must notify the people being processed — employees whose emotional states are being tracked by workplace surveillance AI, for example, or users of apps that categorise mood from facial expressions.

This provision catches more systems than you might expect. AI tools that analyse employee video calls for engagement signals. Customer analytics tools that infer mood or intent from voice tone. Retail AI that estimates age or sentiment from camera feeds. All of these require clear disclosure to the people being processed.

Unlike deepfakes, there’s no creative-use exception here. If the system is categorising or inferring emotional states using biometric data, it must disclose. Full stop.

What About UK Businesses After Brexit?

This is the question I get most from UK readers, and the honest answer is: it’s complicated, but don’t assume you’re exempt.

The EU AI Act applies to any organisation that provides AI systems to users in the EU, or deploys AI systems in the EU market — regardless of where that organisation is headquartered. A UK company running a chatbot accessible to users in France, Germany, or Spain is deploying that chatbot in the EU market. Article 50 applies.

The threshold isn’t having an EU office or paying EU taxes. It’s whether your AI system can be accessed by EU residents. That’s a very broad net, and most UK businesses with any European customer base are inside it.

The UK’s own AI regulation is moving on a different timeline. As of August 2026, there is no UK AI Act equivalent with binding horizontal obligations. The FCA has been moving on sector-specific AI guidance for financial services, and the Information Commissioner’s Office has updated its guidance on AI and data protection. But the UK’s approach remains principles-based rather than the EU’s rule-based model. If you serve EU customers, you’re operating under the EU framework whether or not the UK has equivalent law.

The Fines — and Why They’re Serious

Violations of Article 50 can result in fines of up to €15 million, or 3% of total annual worldwide turnover — whichever is higher. For a large multinational, 3% of global revenue is the bigger number by a long distance. These are not token penalties. They’re calibrated to actually change behaviour at scale.

Enforcement rests with national regulators in each EU member state, plus the European AI Office for general-purpose AI models. The first enforcement actions under Article 50 have not yet been announced — but given how recently the rules came into force, that will change. Early cases are likely to focus on the most visible violations: chatbots that clearly don’t disclose their AI status, and deepfake content that circulates without labelling.

The compliance gap in the market right now is real. Many businesses haven’t updated their chatbots, haven’t introduced deepfake labelling workflows, and are assuming the grace period covers more than it does. That assumption is going to be tested when the first enforcement actions land.

What This Means for UK Readers

In the short term, if you run a UK business with EU-facing customers and use AI chatbots or generative AI tools, check your disclosure practices now. The chatbot disclosure requirement has no grace period. If your customer service bot doesn’t tell EU users it’s AI, you’re already out of compliance.

For consumers, the change is more subtle but real. Every chatbot you interact with on a European website, every AI voice system you call, and every deepfake video shared publicly in EU-accessible media is now legally required to identify itself. It won’t always be obvious — the label might be a small notice at the start of a chat, or a tag in the metadata of a video — but it should be there.

The bigger picture worth watching is the long-term convergence between EU and UK AI regulation. The UK government has signalled it wants a “pro-innovation” regulatory approach that avoids the most prescriptive elements of the EU Act. But UK businesses operating in Europe will need to comply with EU rules regardless. And there is significant political pressure — particularly from consumer groups and civil society — to adopt equivalent transparency rules domestically. Whether the UK follows the EU on AI transparency, as it eventually did on GDPR, is one of the more consequential regulatory questions of the next two years.

This article is for educational purposes only and does not constitute financial advice. Cryptocurrency investments involve significant risk. Always do your own research.