Reply to thread

Message: [QUOTE="Bombastus, post: 91177, member: 2178"] Since August 2, 2026, Article 50(1) has required direct-interaction AI systems to tell people they are dealing with AI from the first exchange. A phone agent does not escape because its voice is generic, licensed, or unrelated to any real speaker. This conversational duty sits beside [B][URL='https://goldmidi.com/community/threads/the-eu-labeling-rules-that-now-apply-to-ai-audio.76573/']the EU transparency framework for synthetic audio[/URL][/B] and does not depend on deepfake status. A plain text-to-speech receptionist can trigger it simply by holding a direct two-way conversation. The formal design obligation falls on the system's provider, which may be the vendor or a business offering the agent under its own name. A compliance promise in a sales contract proves little if the configured call route fails to deliver the notice. [HEADING=2]Direct conversation triggers the disclosure duty[/HEADING] Four conditions define the covered interaction. The product must qualify as an AI system, support a genuine two-way exchange, communicate without a human intermediary, and deal with a natural person. Customers, employees, contractors, and professionals all count as natural persons here. A background speech analyzer that sends suggestions to a human operator does not itself speak to the caller. That mediated arrangement differs from an AI receptionist that hears open speech, decides what to say, and answers in real time. Basic phone automation can sit outside the rule as well. A menu that only records an account number or plays fixed responses may merely collect data, and it may not qualify as AI at all. Once the system interprets free speech and conducts the exchange itself, the direct-interaction test is much easier to satisfy. Outbound agents add another identity problem. When an agent acts for someone else, the guidance expects it to disclose its AI nature and identify the person or organization it represents. If the provider cannot predict whether the recipient is human, it should use [B]an unmistakable spoken AI introduction[/B] whenever human contact is likely. [HEADING=2]The opening notice must survive the call flow[/HEADING] The notice belongs at the start of each person's first interaction, before the agent begins verification, gathers details, or handles the task. “I'm an AI assistant calling for Northstar Bank” states both facts plainly. “This call may be recorded” does not identify artificial intelligence and cannot do the same job. Terms and conditions, a privacy page, or a machine-readable signal cannot replace the spoken notice on a phone call. The information must be clear, distinguishable, and accessible in the medium the person is using. A chime or futuristic voice effect may support the wording, but neither reliably tells a caller that the speaker is AI. Real call flows create the failures that polished demos hide. Barge-in can cut off the greeting, language routing can load a different prompt, and a callback path can bypass the original opening. A disclosure mechanism therefore has to survive retries, transfers, queue changes, and configuration updates rather than living in one happy-path script. Each new listener gets a separate first interaction. If the agent is transferred to a colleague who never heard the opening, the earlier notice does not cover that person. If a human employee starts the call and an AI agent later takes over, the AI must identify itself when its direct exchange begins. Testing the audible result matters more than checking that disclosure text exists in a prompt file. Session records can show whether the notice played, whether the caller interrupted it, and which version ran. They also expose silent failures after a vendor update or routing change. [HEADING=2]The obvious-AI exception remains deliberately narrow[/HEADING] Article 50(1) allows no notice where the AI interaction is already obvious, but the guidance treats that exception restrictively. The reference point is a reasonably well-informed, observant, and circumspect member of the intended audience, considered in the actual circumstances of use. A robotic voice may help, yet vocal style alone is a fragile basis for obviousness. Phone audio removes visual badges, callers join with different levels of technical knowledge, and modern synthetic voices can change without warning after an update. A human name or conversational pause makes the artificial nature less apparent, not more. One prominent opening notice will usually be enough in an ordinary service call. Longer exchanges involving vulnerable people, emotional attachment, health, finance, insurance, legal help, or complaint handling may need periodic contextual reminders. The law sets no universal interval, so frequency should reflect duration, sensitivity, and the risk of forgotten AI identity. The disclosure does not need a legal recital or technical account of the model. It needs plain language that a caller can perceive and understand. An agent with a natural voice and human name should identify itself as AI and name its principal before requesting information, consent, payment, or agreement. [/QUOTE]

Name