关键词: 生成式人工智能, 信息茧房, 认知安全, AI对齐, 对话风险, 可解释评估

Abstract: This paper proposes GCI-MVP,a theory-driven multidimensional diagnostic model that quantifies information cocoon risks in Generative AI dialogues.Unlike traditional recommender systems,Generative AI exhibits “unidirectional compliance”.This tendency risks constructing closed cognitive spaces through human-AI co-construction－manifested as topic narrowing,opi-nion homogenization,and cognitive frame repetition.Existing studies remain largely theoretical and lack fine-grained,interpretable diagnostic tools for dynamic conversational flows.GCI-MVP addresses this gap through three synergistic innovations.Firstly,it establishes a theory-computable-explainable triadic paradigm.This paradigm encodes three core information cocoon mechanisms—cognitive selection bias,machine cognitive resonance,and cognitive frame repetition—into learnable neural computation units:to-pic prototypes,semantic anchors,and frame probes.Secondly,it synthesizes three interpretable metrics—topic cocoon index(TCI),semantic uniformity index(SUI),and frame repetition index(FRI)－via a multi-branch diagnostic architecture.Thirdly,it enables end-to-end risk assessment through a theory-driven linear fusion layer.Experiments on real-world Chinese dialogues(WildChat) demonstrate that GCI-MVP effectively diagnoses information cocoon risks at different levels,achieving 85.4% accuracy and a macro F1-score of 0.825 in three-level risk classification.It significantly outperforms LDA-based topic diversity,lexical diversity,and fine-tuned BERT baselines.Bootstrap tests and McNemar's test jointly confirm the statistical significance of this advantage(p<0.001).Systematic ablation studies validate the necessity of each diagnostic dimension,and typical case analyses further reveal strong alignment between the proposed metrics and theoretical mechanisms such as “cognitive selection bias” and “machine cognitive resonance.” Cross-model generalization tests on GLM-4,ERNIE 4.0,and BLOOMZ-7B demonstrate that the model achieves stable diagnosis without fine-tuning(macro F1=0.803),exhibiting strong generalizability.GCI-MVP provides a computable,interpretable,and auditable tool for assessing cognitive safety risks in generative AI,offering important theoretical value and application prospects for building secure and trustworthy human-AI dialogue systems.

Key words: Generative AI, Information cocoons, Cognitive safety, AI alignment, Conversational risk, Interpretable evaluation