Reply to thread

Message: [QUOTE="Bombastus, post: 91176, member: 2178"] Article 3(60) treats synthetic audio as a deepfake only when three cumulative conditions are met, not merely because AI produced the voice. The output must closely resemble a real or plausibly real subject and be capable of falsely seeming authentic or truthful. That rule separates synthetic production from public deception. Article 50(2) generally makes the system provider responsible for machine-readable marking of synthetic output. Article 50(4) gives the deployer a listener-facing duty only when the finished audio qualifies as a deepfake. Classification turns on the recording as people actually encounter it. The model name, preset label, and producer's private intention cannot settle what a foreseeable listener is likely to believe. Failure on any one condition defeats the deepfake classification, although the provider's separate marking duty may remain. [HEADING=2]Resemblance alone does not complete the test[/HEADING] A famous actor's close clone is the easy example. Yet an identifiable living person is not the legal boundary. Audio can resemble a person, object, place, entity, or event, so a fabricated company announcement or simulated historical broadcast may also need scrutiny. The word existing is wider than it looks. The simulated subject may exist now, may plausibly exist, or could plausibly have existed, which means a realistic fictional speaker is not automatically outside the definition. Nor does calling a voice generic create a safe harbor. A stock narrator with no claimed identity may fail the resemblance test, yet its presentation can supply a person or entity to imitate. Account names, captions, logos, delivery style, and the script itself may make listeners attribute ordinary text-to-speech to a named executive, broadcaster, or public body. High similarity still matters. A shared accent or speaking rhythm may not show resemblance, while copied cadence, characteristic phrasing, and vocal profile can build [B]a convincingly borrowed speaker identity[/B]. You assess the combined impression rather than hunting for one decisive acoustic trait. Article 3(60) sets no numeric similarity score. The inquiry remains contextual, so an internal voiceprint result cannot replace the legal assessment of resemblance and listener belief. [HEADING=2]Audience context decides whether authenticity is false[/HEADING] The same synthetic recording can land differently in two settings. Listeners expect an artificial performance in a voice-tool demonstration. A clip posted as a leaked board call instead invites them to treat its speaker and occasion as real. The third criterion focuses on the intended and reasonably foreseeable audience, not the least attentive person imaginable. Their composition, knowledge, and expectations matter alongside the substantive message and deployment setting. The finished presentation must have the capacity to mislead those listeners about authenticity or truthfulness. The script does not need to contain a factual lie. A cloned chief executive could read accurate public figures and still create the false impression that the executive personally made that recording. Article 50 is concerned with an artificial origin presented as authentic as well as a false substantive message. Intent to deceive is not required either. A producer may be acting carefully, using an approved script and a licensed vocal model, but those facts do not change the audience-facing impression. A licensed clone can therefore fall within [B][URL='https://goldmidi.com/community/threads/the-eu-labeling-rules-that-now-apply-to-ai-audio.76573/']the disclosure duties for AI-made audio[/URL][/B] when listeners could wrongly believe the named person delivered the performance. If all three criteria are met, the deployer must make the artificial origin clear by the listener's first exposure. A provider's hidden or machine-readable mark cannot replace that human-perceivable disclosure, because the two duties serve different recipients. [HEADING=2]Creative framing changes the treatment, not the test[/HEADING] An evidently artistic, creative, satirical, fictional, or similar work gets a lighter disclosure method, not a blanket exemption. The notice may be presented in a way that does not hamper enjoyment, but a qualifying deepfake still needs one. Sequence matters here. First decide whether the audio has high resemblance, a real or plausibly real subject, and a false appearance of authenticity in its context. Only then decide whether creative framing permits a less intrusive disclosure, since a fiction label cannot rescue a performance presented as genuine evidence. A plainly artificial character may fall outside the definition because the audience never expects authenticity. A natural-sounding fictional podcast host is less certain, particularly when the surrounding branding presents that host as an actual person with a real biography. The distinction comes from the audience's expected reading, not from the creator ticking a fiction box behind the scenes. Keep consent and ownership review on a separate track. Unauthorized cloning may create privacy, personality, copyright, contract, or consumer-law trouble even when Article 50's deepfake test fails. A fully licensed clone may still require disclosure, because permission does not change what the audience believes. [/QUOTE]

Name