AI voices and voice clones.

Is it a deepfake? Audio examples

For the AI Act, “persons” include personal characteristics such as a voice, and it is enough that the voice resembles someone who exists or could plausibly exist. What decides is whether listeners would be misled into taking it as a real person speaking. The assessment considers the audience you can reasonably expect, including children or older listeners.

How an audible disclosure works

• An audible disclaimer at the latest at the first exposure, for example at the start of the clip or call (Code of Practice, Section 2, Sub-measure 1.2.3, for signatories).
• For long or live audio, reminders at regular intervals, and at least after interruptions such as adverts; tones or earcons may support recognition.
• For evidently artistic or fictional audio offered in an app or on a website, the Code (Commitment 3) allows the disclosure next to the player, as long as people see it without extra clicks.
• Clips get cut and re-shared: a disclosure only at the very start can be lost, which is why the Code asks for repeats.

Labelling is not consent

Article 50 only says when to disclose. It does not make cloning someone’s voice lawful; the guidelines stress that rights of third parties, such as data protection and intellectual property, must still be respected. From 2 December 2026, Article 5 also bans AI systems used to generate realistic audio, images or video of an identifiable person’s intimate parts or sexually explicit activities without their explicit consent.

A voice agent that talks to people is a second case: it interacts with them, so its provider must tell them they are dealing with an AI (Article 50(1)).

Platform rules for AI voices