Generative AI and Linguistic Mediation: Relational Knowledge, Expression Dynamics, and Provenance

Generative AI increasingly learns from human traces while also entering the processes through which people write, translate, summarize, remember, and represent one another. This paper develops a preliminary framework for evaluating that dual position. Its philosophical and interdisciplinary review connects arguments about syntax and understanding, symbol grounding, tacit and situated knowledge, distributed cognition, multimodal and embodied systems, AI-mediated communication, predictive writing, linguistic inequality, participatory language technology, dataset documentation, and content provenance. The central proposal is a layered warrant profile that separates artifact performance, participation in a source relation, practical successor competence, interpretive standing, expression distribution, individual accessibility, semantic and relational-historical fidelity, provenance, institutional authority, and downstream generativity. A typed mediation chain and an exposure–acceptance–editing kernel locate human and institutional action within the process. An exact finite construction then shows that AI assistance can expand one user's effective expression repertoire while increasing concentration in the population's accepted output distribution. A state-dependent recurrence separates human accommodation, system updates, institutional selection, and training-data feedback; local convergence follows under an explicit contraction condition. A second exact construction establishes that a final artifact alone generally underidentifies its human and AI transformation history. Three mechanism cases involving professional rewriting, minoritized-language assistance, and institutional summaries of shared experience illustrate the framework. The paper concludes with an empirical programme and a governance portfolio joining usable disclosure, selective provenance, privacy, contestation, community authority, portability, and longitudinal review. The resulting position is conservative and revisable: successful artifacts, relational competence, expression change, provenance, and legitimate authority each require evidence suited to their own object.