Beyond Exposure: Persistence, Transfer, and Control Sensitivity in Longitudinal Human–LLM Interaction

Beyond Exposure: Persistence, Transfer, and Control Sensitivity in Longitudinal Human–LLM Interaction LLM-L01 Research Synthesis | Version 0.5 | November 2026 Longitudinal studies of human–large language model (LLM) interaction increasingly report beneficial, adverse, and null outcomes across mental health, companionship, education, workplace performance, and epistemic behavior. However, an important methodological distinction is frequently overlooked: the passage of time does not, by itself, establish that an observed effect persists independently of continued AI access or transfers to unaided human performance. This working paper introduces an operational evidence-classification framework designed to distinguish these outcomes more precisely. The study synthesizes 99 outcome-level records from 14 indexed sources (13 primary empirical studies and one overlapping meta-analysis). Each outcome is classified using an E0–E12 access-state taxonomy that distinguishes on-treatment effects, follow-up assessments with optional or unreported AI access, immediate and delayed unaided testing, independently verified withdrawal, nonrandom interruptions, and transfer to broader tasks. Principal Findings The evidence audit identifies several important boundaries in the existing literature: 14 outcome records from three studies provide evidence of unaided assessment, interrupted AI access, or broader-task transfer (E4–E8), exclusively within educational and workplace settings. No study in the examined corpus independently verifies complete AI non-use throughout a withdrawal interval (E6 = 0). 13 mental-health follow-up records involve optional or unreported AI access, preventing their classification as verified post-withdrawal persistence. The examined social and psychosocial studies contain no verified unaided withdrawal-persistence outcomes. Beneficial, adverse, and statistically nonsignificant results vary across outcome measures, comparison conditions, and assessment schedules. These findings identify a limitation in the available evidence, not proof that lasting beneficial or harmful effects cannot occur. Methodological Contribution The paper provides an auditable framework for distinguishing immediate performance from retained capability, continued-treatment effects from withdrawal effects, and randomized intervention assignment from self-selected engagement. An accompanying versioned evidence register (v0.5) contains the outcome-level extractions, source information, comparison and design classifications, access-state variables, uncertainty estimates, and reproducibility checks. All 99 assigned evidence codes are reproducible from the recorded classification fields, with zero detected coding mismatches. The synthesis also proposes falsifiable directions for future human longitudinal experiments, including independent skill transfer, emotional benefit versus social connection, comparator sensitivity, and reciprocal sycophancy feedback. A separate research protocol, LLM-L02, addresses a candidate model-side mechanism involving persistent memory, retrieval, and evidence-based correction fidelity. Its experimental claims are independent of the human evidence synthesis. Scope and Limitations This report is a structured, purposive, non-systematic research synthesis, not a completed systematic review, new meta-analysis, or experimental demonstration of causal effects. The source corpus is not exhaustive. Some supplementary verification, risk-of-bias assessments, and independent extraction checks remain outstanding. Outcome records within individual studies are not statistically independent. The primary contribution is therefore an operational and reproducible evidence-coverage framework that clarifies what existing longitudinal human–AI studies establish, what remains unidentified, and which experimental designs are necessary to distinguish persistent human adaptation from continued AI assistance. Research status: Working manuscript, version 0.5. Not peer-reviewed. Associated materials: LLM-L01 synthesis manuscript (PDF) and outcome-level evidence register v0.5 (XLSX).