The 'We Disclosed It in the Terms' Myth: Article 50 Says No

"Our chatbot says it's an AI in the terms and conditions, so Article 50 is handled."

No. Buried disclosure is decoration. The EU AI Act requires people to know they're talking to a machine at the moment of interaction — clear, distinguishable, accessible. A legal footer link fails the test even when technically present.

And the regulators' entry point isn't an audit visit. It's this:

┌─────────────────────────────────────┐
│ HOW AN ART. 50 CASE ACTUALLY STARTS │
├─────────────────────────────────────┤
│ 1. Customer complains they were     │
│    "tricked by a fake human"        │
│ 2. Attaches the transcript          │
│ 3. Transcript is the evidence file  │
│ 4. Market surveillance takes it     │
│    from there                       │
└─────────────────────────────────────┘

The nuance most teams miss: it's not only chatbots. A fully autonomous AI that drafts AND sends customer replies — with no human review — puts you in Article 50 territory, because the recipient believes a person answered them. (An assistant where a human sends and owns the message does not.)

The same trap in reverse: relying on your chatbot vendor's watermarking does not discharge your deployer duty. The Commission confirmed the disclosure must be perceivable by a person, without specialist tools.

The safe pattern for automated sends is boring and works:

✓ Disclosure line in the message: "This reply was prepared by our AI assistant" ✓ Header disclosure in widgets ✓ A documented decision trail: what you disclose, where, since when

Because like Article 4, what counts as a measure is what leaves a trace.

60-minute self-audit + the Art. 50(2) synthetic-content marking deadline (2 Dec 2026 for legacy systems): Article 50 guide