If you’ve talked to a customer-service chatbot, seen an AI-generated image on your feed, or watched a synthetic video clip in the last few days, something changed under the hood on August 2, 2026 — even if nobody told you. That’s the day Article 50 of the EU AI Act, the bloc’s transparency rulebook for artificial intelligence, went into full effect. It doesn’t ban anything. It just makes AI systems tell you what they are.
Here’s what actually changed, who has to comply, and whether it matters if you don’t live anywhere near Brussels.
What You’ll Learn
This guide breaks down the four things Article 50 actually requires, who’s exempt, whether it reaches companies outside Europe, and what happens if a company just ignores it.
What Article 50 Actually Requires
The European Commission adopted formal implementing guidelines on July 20, 2026, just under two weeks before the law took effect, giving companies a fairly narrow runway to comply. The rule breaks down into four distinct obligations, and they don’t all apply to the same products:
- Direct interaction with AI. If you’re talking to a chatbot, virtual assistant, or any system designed to simulate a conversation with a human, the provider has to tell you that up front — not three menus deep in a settings page.
- AI-generated content. Images, audio, video, or text produced or meaningfully altered by AI has to be marked as such, typically through visible labels or embedded metadata.
- Emotion recognition and biometric categorization. If a system is inferring your mood, age, or other biometric-adjacent traits from your face, voice, or behavior, you have to be told that’s happening.
- Deepfakes and AI text on public-interest topics. Synthetic media that looks or sounds like a real person, and AI-written text published on matters of public interest (news, politics, elections), both require disclosure.
The Chatbot Rule — What Changes When You Talk to AI
This is the one most people will actually notice. Under the new guidance, a chatbot has to disclose that it’s an AI “clearly and upfront” — at the moment you start interacting with it, not buried in a terms-of-service link you’ll never click. Expect to see more of the “Hi, I’m an AI assistant” openers that some services already use, except now it’s mandatory across the EU rather than a nice-to-have.
The exception is narrow: if it’s genuinely obvious from context that you’re dealing with AI — think an obviously robotic IVR phone tree — the explicit disclosure requirement loosens a bit. But the default assumption in the guidelines is that companies should over-disclose, not under-disclose.
The Deepfake Rule — Even Without Intent to Deceive
This is the part with real teeth, and it surprised a lot of compliance teams: the labeling requirement for deepfakes applies even if nobody was trying to trick anyone, and even if the content doesn’t depict a real person at all. A synthetic video that merely looks photorealistic and could pass for authentic footage can trigger the rule. Intent isn’t the test — realism is.
For AI-generated text specifically, the disclosure requirement is scoped to public-interest content — meaning news, current-events commentary, and similar material, rather than every AI-assisted email or product description you’ve ever sent.
Does This Apply Outside the EU?
Yes, in the same way GDPR reached far beyond Europe’s borders. Article 50 applies based on where the users are, not where the company is headquartered. If your chatbot, app, or AI-generated content reaches people physically located in the EU, the obligations apply to you regardless of whether your servers — or your entire company — are based in Seattle, Bangalore, or anywhere else. Global platforms with EU users are, in practice, complying everywhere at once rather than building an EU-only version of their product.
Exemptions — What’s NOT Covered
The Commission’s guidance carves out four notable exceptions:
- AI systems authorized by law to detect, prevent, investigate, or prosecute crime — as long as safeguards protect other people’s rights, and the tool isn’t a public-facing crime-reporting system.
- Situations where the AI-generated nature is already “obvious from context.”
- Minor assistive editing (think: AI-powered spell-check or a touch-up filter), which doesn’t rise to the level of “AI-generated content” the rule is targeting.
- Clearly artistic, fictional, or satirical works — a deepfake used in an obvious parody sketch isn’t the same problem the rule was written to solve.
Penalties If Companies Ignore It
Non-compliance with Article 50 carries fines of up to €15 million or 3% of a company’s total worldwide annual turnover, whichever number is bigger — the same penalty tier used for several other AI Act violations. Enforcement isn’t centralized in Brussels; each of the 27 EU member states designates its own national market surveillance authority to investigate and issue fines, similar to how GDPR enforcement is split across national data protection authorities.
Pro Tip: How to Actually Verify a Disclosure
If you want to check whether a chatbot or piece of content you’re looking at is properly disclosed, look for it at the point of first contact — a chat window’s opening message, a small “AI-generated” badge on an image, or metadata visible when you check a file’s properties. If the only place a company mentions AI involvement is a general privacy policy, that’s exactly the kind of buried disclosure the new guidelines were written to eliminate.
This isn’t the EU’s only recent swing at reining in how tech companies treat users — we covered Europe’s crackdown on Meta’s “addictive design” practices in a similar vein, and our recap of the White House’s AI safety meeting with OpenAI, Anthropic, and Google shows how differently the US is approaching the same underlying problem. If you’re trying to figure out which AI products are worth trusting with your data in the first place, our guide on which AI models are actually worth switching to is a good next stop — and if you’ve ever wondered whether a glowing product review was written by a bot, our guide to spotting fake AI-generated reviews covers the exact same disclosure gap from the buyer’s side.
Frequently Asked Questions
Does Article 50 apply to me if I’m not in the EU?
Only if you’re interacting with a company’s AI product from inside the EU. But because most global companies build one compliant version of their product rather than a separate EU-only build, you’ll likely see the same disclosures no matter where you are.
Does this mean AI chatbots are banned in Europe?
No. Article 50 doesn’t restrict what AI can do — it only requires disclosure. A chatbot is fully legal; it just has to identify itself as AI.
What counts as a “deepfake” under the law?
Any audio, image, or video content that resembles a real existing person, place, or event closely enough that it could be mistaken for authentic, whether or not the creator intended to deceive anyone.
Who actually enforces these rules?
Each of the 27 EU member states has its own national market surveillance authority responsible for investigating violations and issuing fines — there’s no single EU-wide enforcement body.
Bottom Line
Article 50 won’t change what AI tools can do, but it will change how honest they have to be about doing it. If a chatbot, image generator, or content platform you use doesn’t tell you plainly when AI is involved, that’s no longer just a trust issue — as of August 2, 2026, it’s a compliance issue in the EU, with an eight-figure price tag attached for companies that ignore it.
Leave a Reply