MetaCity Has Been Cloning User Voice Profiles for AI Training Since 2024 — The Opt-Out Setting Is Buried Six Menus Deep, Labeled 'Audio Personalization Enhancement,' and Defaults to On for All Accounts Created Before March 2025

A data privacy researcher published findings this morning showing that MetaCity's voice cloning data collection program — which captures and processes vocal audio from all live streams, voice messages, and avatar speech interactions — has been active since September 2024. The opt-out mechanism exists but requires navigating six nested settings menus and is labeled 'Audio Personalization Enhancement' with no mention of AI training or voice cloning. The setting defaults to enabled for all 340 million accounts created before March 2025, when a ToS update quietly introduced the program. MetaCity's Data Policy page has not been updated to reflect the practice.

Dr. Sable Wen spent three weeks on this. The research methodology, published alongside the findings at 7:30 AM, documents a systematic audit of MetaCity's Terms of Service amendments from 2024 to present, a full mapping of the platform's privacy settings architecture, API traffic analysis during voice-active sessions, and a line-by-line comparison of the March 2025 ToS update against the prior version. The key finding — that MetaCity has been collecting and processing user vocal audio for AI voice model training since September 2024 — emerges from the intersection of three data points: a clause in the March 2025 ToS update that grants MetaCity 'a perpetual license to use audio data derived from user interactions for platform improvement and AI model development,' API traffic logs showing audio data transmission to a separate processing endpoint during live voice sessions, and a settings toggle that, when disabled, stops that transmission. The toggle is six menus deep. It is not labeled with anything that would cause a user to associate it with AI voice training.

The settings path requires documentation because 'six menus deep' is an accurate description that understates the navigational intent. To reach the Audio Personalization Enhancement toggle, a user must: open Settings from the main account menu (menu 1), navigate to Account from the Settings categories (menu 2), select Privacy from the Account subcategories (menu 3), choose Data & Personalization from the Privacy options (menu 4), open Audio Services from the Data & Personalization subsection (menu 5), and locate the Audio Personalization Enhancement toggle within Audio Services (menu 6). The toggle is the third of seven options in that menu. Its description text reads: 'Allow MetaCity to use your audio interactions to personalize and improve your experience.' The words 'AI training,' 'voice cloning,' 'model development,' and 'data licensing' do not appear in the description. Dr. Wen's report notes that this architecture 'appears consistent with a deliberate design choice to minimize user discovery of the opt-out pathway.'

Your Voice Has Been Collected. The Setting to Stop It Is Where They Knew You Wouldn't Look.

The March 2025 ToS amendment is the document that made the program legally operational, and the amendment's construction is worth examining in detail. The relevant clause — section 14(c) in the updated ToS — appears between two uncontroversial provisions about content archiving and platform analytics. It reads: 'By using MetaCity's audio-enabled features, you grant MetaCity a perpetual, worldwide, royalty-free license to collect, process, store, and utilize audio data derived from your platform interactions for the purposes of platform optimization, feature development, and AI model training, subject to your privacy settings.' The phrase 'subject to your privacy settings' is the clause that creates the opt-out — but only if a user knows the opt-out exists, knows where to find it, and understands that 'Audio Personalization Enhancement' in a submenu is the mechanism by which they exercise it. Dr. Wen's survey of 200 MetaCity users found that none could locate the toggle without guidance, and none associated the toggle's label with voice cloning.

The scale of the data collection is what distinguishes this from a standard dark-pattern privacy complaint. MetaCity has approximately 340 million accounts created before March 2025, all of which have been opted in since September 2024. An active MetaCity user who streams twice a week for an hour each session has contributed approximately 150 hours of vocal audio to the program over 18 months. Creators — who stream for longer and more frequently — have contributed significantly more. The total estimated pool of collected voice data, assuming even modest per-user engagement rates, represents one of the largest privately held vocal training datasets in existence. What voice models that data has trained, who those models have been licensed to, and what those models are capable of producing are questions the published findings raise but cannot answer, because MetaCity has not disclosed the downstream use of the collected audio.

The community response has two immediate practical dimensions. The first is navigational: Dr. Wen's report includes a step-by-step guide to reaching the opt-out toggle, and that guide is currently being shared at a rate that suggests a significant portion of MetaCity's active user base is in the process of disabling the setting. The second is legal: four separate data protection advocacy organizations in EU jurisdictions have posted statements since 7:30 AM indicating they are reviewing the findings for potential regulatory action. The EU Digital Wellbeing Directive, already invoked this morning in connection with the PREEMPT disclosure, also contains provisions covering undisclosed AI training data collection from behavioral signals — provisions that several legal commentators believe apply to the voice cloning program as described. MetaCity's data policy page, last updated in February 2025, contains no reference to voice cloning, AI voice model training, or the Audio Personalization Enhancement setting. It has not been updated today.