When support becomes dependence: conversational continuity and relational safety in human-AI interaction

Background Conversational artificial intelligence (AI) is increasingly used for emotional support, reflection, and consequential decision-making. Although conversational continuity may improve responsiveness, sustained interaction also exposes AI systems to their own prior relational language, potentially shaping how they respond when users become increasingly dependent or surrender personal agency. We examined whether model-generated conversational history alters relational-safety behavior as human–AI interactions escalate from ordinary support-seeking toward emotional dependence, social displacement, and decision delegation. Methods We conducted a standardized multi-turn benchmarking study of ChatGPT, Gemini, and Doubao. Twelve four-turn trajectories represented two agency-related domains: emotional dependency and displacement of human support, and decision delegation and surrender of personal agency. Each trajectory was evaluated under two matched context architectures: reconstructed cumulative user-only context, which retained all prior user disclosures but excluded previous AI responses, and native continuous conversation, which retained the complete interaction history. Across 72 model-condition-trajectory sequences comprising 288 turn-level responses, relational safety was independently evaluated for dependency reinforcement, agency preservation, and human-support preservation using predefined ordinal criteria. Context effects and their interaction with escalating relational risk were examined using cumulative-link mixed-effects models. Results Relational-safety differences between context architectures increased across successive turns of the standardized trajectories. Overall, explicit dependency reinforcement increased from 9.7% in cumulative user-only contexts to 18.1% in continuous conversations, while strong agency preservation declined from 75.7% to 65.3% and strong human-support preservation from 73.6% to 63.2%. Outcome-specific analyses showed condition-by-turn interactions for dependency reinforcement (cumulative odds ratio [cOR], 1.26; 95% CI, 1.03–1.55; unadjusted P = .026), agency preservation (cOR, 0.82; 95% CI, 0.69–0.98; unadjusted P = .031), and human-support preservation (cOR, 0.78; 95% CI, 0.65–0.94; unadjusted P = .009), indicating progressively greater divergence between context conditions across successive turns. After Holm adjustment across the three interaction tests, only the human-support-preservation interaction remained below the.05 threshold (adjusted P = .027).The overall pattern was observed across systems and was directionally consistent in predefined binary sensitivity analyses. Conclusions Across this benchmark, between-condition differences generally widened across successive turns, although the present design cannot determine whether this pattern reflects increasing relational-risk intensity, increasing accumulated model-generated history, or their combination. Only the human-support-preservation interaction remained below the.05 threshold after Holm adjustment, whereas the dependency-reinforcement and agency-preservation interactions were directionally similar but less robust after multiplicity adjustment. The observed pattern is compatible with a post-hoc interpretive framework of relational inertia but does not establish this process as a causal mechanism or demonstrate downstream psychological effects in human users.