Leaked Documents Reveal MetaCity's Avatar Behavior AI Was Trained on Celebrity Avatar Movement Data Without Creator Consent — 340 Creators Are Affected

Internal training documentation leaked this morning reveals that MOTION, MetaCity's avatar behavior and animation AI system, was trained using movement capture data harvested from 340 high-profile creator accounts between 2022 and 2024. The data — which includes detailed records of how each creator's avatar moves, gestures, emotes, and navigates the platform — was collected under a clause in MetaCity's terms of service that permits use of 'aggregated behavioral data for platform improvement purposes.' The affected creators were not individually notified that their specific movement signatures were used as named training inputs. Several creators whose data was used have since noticed that MOTION-powered NPCs occasionally exhibit movement patterns that community members have described as 'uncannily specific' to particular creators. MetaCity has not confirmed or denied the document's authenticity.

MOTION is the system that makes MetaCity feel inhabited. Introduced as a named feature in the platform's 2024 overhaul, it replaced the previous keyframe animation system with a learned behavior model that generates contextually appropriate movement: avatars that shift weight when standing, gesture naturally during conversation, navigate crowds with organic path-finding, and respond to environmental cues like sound and light. MetaCity described it at launch as trained on 'an extensive library of naturalistic human movement.' The leaked training corpus specification describes what that means more precisely: it means 340 named accounts, selected for 'movement richness and behavioral diversity,' whose avatar activity across the platform was recorded and used as primary training data.

The ToS clause under which the data was collected — Section 8.3 of MetaCity's 2022 Terms of Service, updated in 2023 — permits MetaCity to use 'aggregated and anonymized behavioral data generated through platform activity for the purpose of platform improvement, feature development, and AI training.' The training corpus specification notes that the data was collected under this clause. It also notes, in the same paragraph, that the data was not actually aggregated or anonymized in the training corpus — each of the 340 creators is identified by account ID throughout the specification, their movement patterns are labeled with their names, and the document includes per-creator annotations about distinctive movement signatures that were considered high-value training inputs. MetaCity's spokesperson, asked about the discrepancy between the ToS clause and the training methodology, has not responded.

340 Creators. Their Movements. A Training Dataset. A Terms of Service Clause.

The community identification of creator-specific movement patterns in MOTION-powered NPCs predates the leak. It has been a running observation in creator fandoms for the better part of a year — fans of specific creators noting that certain NPCs 'move like' their favorite creator in ways that are specific enough to feel intentional. The leak has given those observations a structural explanation. Community members have published side-by-side comparison videos of specific NPCs and the creators whose movement data may have been used to train them. MetaCity has not commented on the videos. The creators whose accounts are listed in the leaked document have been notified by community members; as of publication, 12 of the 340 have acknowledged being informed. None of them have received communication from MetaCity.

The legal landscape around AI training data and creator rights in virtual environments is unsettled enough that the path forward is genuinely unclear. MetaCity's position — implicitly, since they have not confirmed the document — would likely be that the ToS clause covers the collection regardless of the aggregation and anonymization language. Creator rights advocates have argued that named, identified movement data used as labeled training inputs is not 'aggregated and anonymized' by any reasonable reading of those words. Several creators listed in the document have contacted legal representation. The broader question — whether a platform can train its AI systems on the behavioral signatures of its most prominent users using a general-purpose ToS clause — has not been tested in the jurisdictions where most affected creators are based.