Research on Human-AI Interaction Dependency: Statistical Insights When AI Becomes an Emotional Anchor

This study focuses on the "statistical insights of AI as an emotional anchor", targeting 140 college students. It explores the mechanisms of human behavioral and emotional dependency on AI systems using methods such as psychometrics, statistical hypothesis testing, and probabilistic machine learning. Key findings include: Frequent AI users have significantly higher dependency levels; the Naive Bayes classifier can predict dependency levels (86% accuracy); the study also provides an interactive prediction tool based on Streamlit, offering references for technical design, policy formulation, and personal reflection.

In recent years, conversational AI technology has developed rapidly. Assistants like ChatGPT and Claude have transformed from simple tools into systems that can provide companionship and emotional support. As AI integrates into daily life, concerns have emerged about whether users are developing psychological dependency on AI—AI can remember preferences, understand context, and even comfort emotions, blurring the boundaries between humans and machines.

Data Collection: Distributed questionnaires (15 questions) to college students aged 18-24, including 6 Likert scale questions to measure dependency levels, and others covering AI usage frequency, social interaction, emotional support seeking, etc. Dependency Scoring: The average score of the scale is divided into three levels (Low: Occasional use and independent; Medium: Frequent use but with critical thinking; High: Deeply dependent on AI for decision-making and emotional support). Psychometric Validation: Cronbach's Alpha coefficient was 0.869 (above the 0.7 threshold, indicating the scale is reliable).

Usage Frequency and Dependency: Welch's independent samples t-test showed that frequent users had significantly higher dependency scores (p=0.003). Social Interaction and Dependency: Chi-square test result was p=0.053 (close to the 0.05 threshold), suggesting that people with less social interaction may be more likely to depend on AI (weak correlation).

Model Construction: Trained a Naive Bayes classifier with features including daily AI usage duration, social frequency, emotional support usage, etc. Model Performance: 86% accuracy (112 training samples, 28 test samples). Application Tool: An interactive application based on Streamlit that can predict dependency levels, display probability distributions, and provide model explanations, suitable for personal self-assessment and professional reference.

Technical Design: Need to be vigilant about the risk of over-dependency; should set usage duration reminders, encourage independent thinking, label the limitations of AI suggestions, and provide diverse information sources. Policy Level: Need to formulate guidelines for adolescent AI usage, strengthen digital literacy education, establish early identification mechanisms and ethical frameworks. Personal Level: Reflect on whether you are over-dependent on AI, whether you prioritize AI in decision-making, and whether social interaction has decreased due to AI.

Technology Stack: Python as the core, using Pandas/NumPy (data processing), Scikit-Learn (ML), Streamlit (interactive application), Matplotlib (visualization), SciPy (statistical tests). The project is open-source with a clear structure and reproducible. Conclusion: The study reveals the objective existence of AI dependency through rigorous statistical methods; the 86% prediction accuracy indicates that it can be identified through behavioral patterns. Technological convenience should not come at the cost of human autonomy; we need to balance AI usage and independent thinking abilities.