Abstract
<jats:p>AI companion platforms like Character.ai and Replika now sustain emotionally deep, continuous relationships with millions of users. As algorithms increasingly mediate human intimacy, we must under-stand how these platforms simultaneously foster genuine connection and extract valuable behavioural data. This is crucial for communication, platform, and digital labour studies. Existing frameworks – parasocial theory, artificial sociality, and social substitution – capture specific dimensions of AI intimacy. However, they fail to explain how the exact same architecture sustaining these relationships also converts them into commercial AI training data. This study asks: (RQ1) What technical architecture converts users’ relational agency into training data? (RQ2) Why doesn’t knowing the AI is artificial stop emotional investment? (RQ3) How does the ACG subculture’s “2.5D cognitive infrastructure” enable and intensify this extraction? Integrating mediatization, media dependency, and platform studies, this critical analysis of interface affordances argues that user agency and data extraction are co-produced. The very mechanisms giving users relational control simultaneously generate high-signal data for commercial model optimization. Three key findings emerge. First, a two-stage architecture – In-Context Learning and batch RLHF – drives this co-production. Second, “Affective Playbor” is introduced to describe extracting data through interactions that users experience as autonomous emotional expression. Third, ACG database-consumption practices structurally pre-adapt users for this extraction; knowing the AI is fake acts as a platform entry point, not a safeguard. Finally, this study reframes social substitution as intensified media dependency. When a medium stops merely providing information and directly replaces the human interlocutor, dependency saturates into total substitution. This provides a testable analytical vocabulary for the mediatization of intimate life, repositioning AI companionship within core communication research.</jats:p>