This study contrasts human‑ChatGPT multi‑turn dialogues with human‑human conversations to examine how morality, politeness, and alignment—three core dimensions of cooperative dialogue—behave across the two settings. We gathered 15,881 human‑ChatGPT exchanges and 10,784 human‑human exchanges, applying mixed‑effects models to identify turn‑level predictors of alignment. The findings reveal that AI reproduces the surface cues of cooperative communication but lacks the underlying social architecture. Moral statements appear pre‑configured rather than negotiated; warmth is expressed without face‑sensitivity; linguistic convergence steadily declines as the conversation progresses. Crucially, mechanisms that normally foster human accommodation reverse in AI interaction: hedging and softening, which boost alignment between humans, reduce alignment when produced by AI; framing in terms of purity, which drives divergence among humans, leads users to converge toward the AI. Agency—the degree to which users can shape the exchange—emerges as the most consistent alignment predictor in both interaction types, while newer models’ lower moral assertiveness does not translate into better cooperation. Overall, AI mimics the outward form of cooperation without the mutual adaptation that grounds it among humans, and human accommodation mechanisms can operate oppositely with AI, suggesting that a turn‑level view alone is insufficient for successful interaction.
Review