Reply to thread

Message: [QUOTE="Bombastus, post: 91149, member: 2178"] Article 50 of the EU AI Act started applying on 2 August 2026, and it puts a disclosure duty on anyone publishing synthetic voice audio in Europe. That includes small publishers, not just model builders. The duty splits across two parties, which is where most people get it wrong. The company that built the voice model carries one obligation. You, the person putting the audio in front of listeners, carry a different one. Getting it wrong is expensive. National market surveillance authorities enforce the rules, and penalties reach 15 million euros or 3 percent of worldwide turnover for the previous financial year, whichever is higher. Smaller companies get proportionality, not a pass. [HEADING=2]Your tool's watermark does not cover you[/HEADING] The vendor's job is machine-readable marking. Providers of systems that generate synthetic audio, images, video, or text must make their outputs detectable as artificially generated, using solutions that are effective, interoperable, and robust as far as the technology allows. Your job is different and more visible. If the audio is a deepfake, meaning content that resembles a real person and could be taken as authentic, you have to disclose that it is artificial in a way an ordinary listener can perceive without any technical tools. A watermark buried in the file does not satisfy that. The regulation asks for something a human notices, which in audio means a spoken note or an equivalent label attached to the content rather than hidden inside it. Timing is specified too. The disclosure has to arrive clearly at the latest on first exposure, not halfway through and not in a description nobody opens. It also has to meet accessibility requirements, so a purely visual label on an audio file is a weak answer. Live voice agents carry a separate rule of their own. If your system speaks with people directly, you have to tell them they are dealing with AI, unless that would already be obvious to a reasonably well-informed person in the situation. The test is the listener's reasonable perception rather than your own sense of how clear the setup is. Build the disclosure into the opening line, and the question stops mattering. None of this is free for the vendor either, and compliance engineering is now a permanent line in [B][URL='https://goldmidi.com/community/threads/elevenlabs-and-the-economics-of-ai-voice.76571/']the cost base of a synthetic voice business[/URL][/B]. [HEADING=2]The grace period does not apply to publishers[/HEADING] There is a delay written into the rules, and it is narrower than most people assume. The extra time runs to 2 December 2026 and covers marking and detection under Article 50(2), which is the provider's technical obligation. Your disclosure duty has no such window. It has been live since August, so [B]telling listeners the voice is synthetic[/B] is already required rather than something to schedule for December. Alongside the law sits a voluntary code. The Code of Practice on Transparency of AI-generated Content was finalized on 10 June 2026, and around 190 companies and organizations had signed it by the end of July. Signing helps, but it does not replace anything. The Commission treats the code as an adequate tool for demonstrating compliance, while the underlying Article 50 obligations stay legally binding whether you sign or not. [HEADING=2]Creative work gets a lighter disclosure, not an exemption[/HEADING] Music, fiction, satire and other creative output are handled separately, and this is the part worth reading closely if you publish anything artistic. The rules still require disclosure. They just soften how it appears. For work that is evidently artistic, creative, satirical, or fictional, transparency narrows to disclosing that generated or manipulated content exists, in a manner that does not hamper the display or enjoyment of the work. A note in the credits can be enough. An announcement over the opening bars is not demanded. That carve-out is about presentation, not about escaping the duty. A synthetic vocal on a released track still needs an honest marker somewhere a listener can find it. Editing tools sit outside the marking rule as well. Systems performing assistive editing functions, or ones that do not substantially alter the input data, are exempt from the machine-readable marking obligation entirely. The practical line is whether the tool changed the substance. Cleaning up a take you recorded is treated differently from generating a performance that never happened, and the difference decides which obligations attach. Text follows its own logic. AI-generated or manipulated text published on matters of public interest must be labeled, unless it went through human editorial review and a named person or body holds editorial responsibility for it. Superficial checking does not count. The review has to be a deliberate examination by someone qualified, with genuine authority over what gets published, which rules out a quick skim before hitting post. [/QUOTE]

Name