EU AI Act chatbot disclosure: what you actually have to do by 2 August 2026

If your website runs an AI chatbot that talks to visitors in the EU, Article 50 of the EU AI Act (Regulation (EU) 2024/1689) applies to that interaction from 2 August 2026. The June 2026 "Digital Omnibus" delayed many parts of the Act — high-risk obligations moved to 2027–2028 — but the transparency rules were not delayed. This guide is plain English, based on the official text and the Commission's May 2026 draft guidelines. It is not legal advice.

The rule in one sentence

People must be told they are talking to an AI system — clearly, at the latest at the moment the interaction starts — unless it is already obvious to a reasonably well-informed person (Art. 50(1) and 50(5)).

Who is on the hook?

The design duty sits with the providerof the AI system. But if you deploy a chatbot on your site, don't relax: vendors like Intercom explicitly state in their terms that customersare responsible for informing end users; several ship their AI labels off by default; and if you white-label a bot under your own brand, legal commentators warn you may take on provider duties yourself. In practice: whoever's website the visitor is on carries the reputational and enforcement exposure, and the practical fix is identical either way.

What counts as a valid disclosure

• An explicit statement, e.g. "You are chatting with an AI assistant, not a human."
• Shown at or before the first message — the bot introducing itself as AI is the cleanest pattern.
• Best practice per the draft guidelines: two layers — a persistent visual "AI" indicator near the chat, plus the first-message notice.
• In the visitor's language for multilingual audiences, and perceivable by assistive technology.

What fails (per the Commission's draft guidelines)

• Disclosure buried in Terms & Conditions or a privacy policy.
• A bot name alone ("Assistant", "Emma") — worse with a human name and avatar.
• Tiny, faint, or momentarily flashing text; a "Powered by AI" footer no one sees.
• English-only notices for a multilingual EU audience.
• Relying on the "it's obvious" exception without documenting why — helpdesk chatbots are the guidelines' example of NOT obvious.

Fines

Up to €15,000,000 or 3% of total worldwide annual turnover, whichever is higher (Art. 99(4)(g)) — for SMEs and startups, whichever is lower. Enforcement by national market surveillance authorities starts 2 August 2026. Early enforcement will likely target egregious cases, but procurement questionnaires and client contracts are already asking about AI Act compliance today.

The related duties people forget

Art. 50(2)–(4) add separate obligations: machine-readable marking of AI-generated content (for systems on the market before 2 Aug 2026, that one got a grace period to 2 December 2026), deepfake labeling, and disclosure of emotion-recognition systems. If you publish AI-generated text on matters of public interest without editorial responsibility, disclosure applies there too.

The 10-minute compliance path

1. Scan your site below (free) to see what's detectable.
2. Set your bot's first message to an explicit AI disclosure, in your customers' languages.
3. Turn on your vendor's AI label (Intercom's is off by default for newer workspaces).
4. Add a persistent AI indicator near the chat and a public AI transparency page.
5. Keep evidence: screenshots, dates, wording — so you can prove it later.

Our Compliance Pack generates steps 2–5 for you: disclosure widget in all 24 EU languages, transparency page, vendor-specific setup guide, and an evidence PDF. One-time €29.