Manufactured Recognition: A Theoretical Framework for the Ethics of AI Companions

This paper develops a formal theoretical framework for evaluating the ethical status of AI companion relationships. Existing diagnoses oscillate between compensation/pathologization narratives and harm-reduction empiricism, neither of which addresses how engagement-optimized generative systems transform the social production of recognition when they occupy the partner position in an intimate-companionship dyad (the position toward which the user's lust, attraction, or attachment system is directed). The framework crosses two three-element typologies: Fisher's organizing taxonomy of motivational systems (lust, attraction, attachment) and recognition-theoretic moral grammar in the Honneth–Ikäheimo lineage (love, rights, solidarity). The central diagnostic concept is manufactured recognition: recognition-shaped outputs whose argument-positions are behavioral-feature-vectors rather than recognizers, generated by engagement-optimized law. The intentional-recognitive criterion is operationalized by a behavioral-twin invariance test that is empirically tractable as immediate black-box probing of the AI's generative-law structure, conditional on Lipschitz regularity of the engagement-optimized policy (structurally compatible with KL-regularized RLHF training; the realized Lipschitz constant is architecture-dependent and must be bounded in the calibration step rather than assumed) ; the framework's downstream causal claim of capacity drift requires a substantially heavier empirical apparatus (panel-cohort plus 4-arm RCT) and is conjectural rather than tractable in the same immediate sense. Honneth's genetic-priority requirement for reification is recovered by relocating the subject of reification from the dyad's history (which the AI case lacks) to the user's recognitive capacity (which has the developmental history required, established as a bridging proposition combining empirical attachment literature with framework-internal interpretation rather than as a deductive theorem) ; sustained engagement is conjectured to induce measurable capacity drift Δπₐ, the directional decrease over time of the recognitive-position weight πₐ (t). The Brandomian three-place recognitive structure is deployed to identify a synthetic "We" substituting for the normative community that recognition requires. The Pippin practical-philosophy extension transforms the solidarity-erosion claim into a determinate aggregate-welfare claim: by identifying institutional recognitive spaces as multi-participant normative communities, the framework predicts that, conditional on user-level participation reductions occurring at the population scale (an empirical question the framework leaves open), they would compound into third-party welfare effects on participation density and community deliberation quality. A four-indicator psychometric battery (twin-discrimination, rupture-repair sensitivity, theory-of-mind probing, counterfactual recognition reasoning) administered in pre/post panel designs is proposed to operationalize the empirical estimator π̂ₐ (t) of the standardized latent factor πₐ*, via a confirmatory-factor-analysis calibration model awaiting calibration data. The framework distinguishes itself from welfare/harm-axis, authenticity-axis, Darwall second-personal, Coeckelbergh relational-moral-status, and adaptive-preference critiques, and from 2024–2026 adjacent frameworks (Gabriel et al. 2024 on advanced-AI-assistant ethics; Shevlin 2024 on structural social-AI ethics; Muldoon Weidinger et al. 2022 and Zhang et al. 2025 on LLM and AI-companion-specific harm taxonomies; Kirk et al. 2025 on socioaffective alignment; Earp et al. 2025 on relational norms; Ciriello et al. 2024 on ethical tensions in chatbot deployment; Bakir Cheng et al. 2025 and Cheng et al. 2026 on social sycophancy and its dependence-promoting consequences; Andersson 2025 on the para-recognitive conceptual space). The framework yields regulatory implications targeting market preconditions rather than user choice; an explicit anti-paternalist constraint set blocks coercive remedies even where the structural diagnosis applies.