Here's the whole thing in two sentences, because that's what an answer engine should be able to quote:
On August 2, 2026, the EU AI Act's Article 50 transparency rules take effect. If you deploy a chatbot or AI voice agent to users in the EU, you must make sure they know they're talking to an AI — and that duty falls on you, the deployer, not on the model vendor you buy from.
Everything below is the practical version. This is a summary for builders, not legal advice; if you're at real scale or in a regulated vertical, get counsel.
The one rule that touches almost everyone#
The AI Act splits duties between providers (who build/place a model on the market) and deployers (who put an AI system to use in their product). Most solo founders and small teams are deployers: you call an API and ship a chatbot.
Article 50's core deployer-facing rule is simple. An AI system "intended to interact directly with natural persons" must be built and run so that a reasonably informed user understands they're interacting with an AI — communicated clearly, at the latest at the first interaction, in an accessible way. That's it. No form, no wall of legalese. A visible notice at the top of the conversation.
The reason it's a headline is the split: because it's a deployer obligation, your model provider cannot comply for you. OpenAI or Anthropic supplying the model doesn't render the disclosure in your UI. You do.
What to ship this week — the checklist#
- [ ] Add an AI disclosure to every conversational surface serving EU users — web chat, in-app assistant, voice agent, WhatsApp/SMS bot. One plain sentence at the start of the interaction.
- [ ] Make it visible before the user's first message, not buried in a privacy policy. "At the latest at the first interaction" means it can't be a link they never click.
- [ ] Label synthetic media if you generate deepfake-style image, audio, or video: mark it as artificially generated.
- [ ] Label AI-written public-interest text (news-style content informing the public) as AI-generated — unless it went through human editorial review and someone holds responsibility.
- [ ] Decide geo-gating vs. global. A one-line notice is usually cheaper to show everyone than to detect-and-branch by region.
- [ ] Don't panic about watermarking. The machine-readable marking standard for synthetic content has a grace period to December 2, 2026 — the human-facing disclosures are what start August 2.
The disclosure, in code#
For a text chatbot, this is genuinely a one-liner rendered before the first exchange:
<!-- Shown as the first message / above the input, not hidden in a policy page -->
<div class="ai-disclosure" role="note">
You're chatting with an AI assistant. Responses are generated automatically
and may be inaccurate.
</div>
For a voice agent, say it in the greeting:
"Hi — quick heads up, you're speaking with an AI assistant. How can I help?"
For an API-driven surface where you control the system prompt, don't rely on the model to announce itself — the disclosure is a UI element you render deterministically, because a model can be talked out of it. Ship it in the interface, not the prompt.
The exemptions (don't over-comply, but don't over-trust them)#
Two carve-outs matter. First, obvious from context: you don't have to state what a reasonable user already understands. A widget clearly branded "AI Assistant," or a product whose whole premise is that it's an AI, can satisfy the standard. But "obvious to me, the builder" is not the test — "obvious to a first-time user" is, and the safe default is still an explicit notice. Second, law-enforcement uses authorized to detect or investigate crime are carved out — not relevant to most founders.
Why the date has teeth#
August 2 also switches on the European Commission's power to fine general-purpose-AI non-compliance — up to €15M or 3% of global annual turnover, whichever is higher. Transparency breaches sit in a lower penalty tier than deploying a prohibited system, but for a small company "lower tier" is still an existential number. The cheap insurance is the sentence above.
A model can be argued out of admitting it's a model. A UI element can't. Render the disclosure in your interface, not your prompt.
If you ship AI globally, this is also the shape of things to come: it rhymes with China's persona-disclosure rules and slots into the broader split governance world founders now build inside. The specifics will vary by market; the baseline — tell the user it's an AI — is quietly becoming universal. Do the five-minute version today, and you're ahead of the rule instead of behind it.



