Abstract
Excessive dependence on Conversational Artificial Intelligence (CAI) can significantly impact individual adaptation and development. This study presents the development and psychometric validation of the Conversational AI Dependence Scale (CAIDS), a 20-item instrument designed to assess CAI dependence among Chinese college students. Through a multi-stage mixed-methods approach, we identified four core dimensions: uncontrollability, withdrawal symptoms, mood modification, and negative impacts. Validation across multiple samples demonstrated high reliability and robust structural validity. Findings indicate that CAI dependence is a significant predictor of negative psychological outcomes, including depression, anxiety, and sleep disruption, while negatively predicting subjective wellbeing.
Introduction
As Large Language Models (LLMs) like ChatGPT become ubiquitous, the landscape of digital addiction is shifting. Unlike the passive consumption typical of social media, Conversational AI (CAI) facilitates dynamic, personalized, one-on-one interactions. This unique engagement model introduces new psychological risks, specifically "cognitive outsourcing"—the delegation of critical thinking and emotional processing to AI agents. Existing tools for measuring Problematic Internet Use (PIU) fail to capture these nuances, necessitating a specialized instrument to quantify this emerging behavioral dependency among digital natives.
Methodology
The researchers employed a three-study design to construct and validate the CAIDS:
- Construct Identification & Scale Development: Semi-structured interviews with high-frequency users identified core themes. An initial 51-item pool was refined to 20 items using Exploratory Factor Analysis (EFA) on a sample of 547 students.
- Structural Validation: Confirmatory Factor Analysis (CFA) was performed on a second sample (n=687) to verify the four-factor structure and internal consistency (Cronbach’s alpha = 0.94).
- Status Survey & Correlation Analysis: A broad survey of nearly 2,000 students mapped the prevalence of dependence and its relationship with mental health indicators using the DASS-21 and the Athens Insomnia Scale.
The Four Dimensions of CAI Dependence
The CAIDS-20 categorizes dependence into four distinct psychological dimensions. Exploratory factor analysis confirmed that these dimensions explain a significant portion of the variance in user behavior, with high factor loadings across all 20 items.
- Uncontrollability: The inability to regulate the time spent interacting with the AI.
- Withdrawal Symptoms: Experiencing irritability, anxiety, or distress when the AI is inaccessible.
- Mood Modification: Utilizing the AI as an emotional crutch to escape negative affect or stress.
- Negative Impact: Real-world consequences, including impaired social functioning and "cognitive laziness."
Validation and Reliability
The scale demonstrated strong discriminant and convergent validity. Correlation analysis showed that while the four dimensions are related, they represent distinct facets of the dependency construct. The square root of the Average Variance Extracted (AVE) for each dimension exceeded its correlation with other dimensions, confirming robust psychometric properties.
Key Findings: Mental Health and Demographic Trends
The study revealed that CAI dependence is a powerful predictor of maladaptive psychological states. Total CAIDS scores showed strong positive correlations with depression (β = 0.53), anxiety (β = 0.56), and friendship difficulties (β = 0.74). Furthermore, high dependence significantly predicted poor sleep quality (β = 0.45), likely due to the cognitive stimulation inherent in interactive dialogue.
Demographically, dependence levels were significantly higher in male students and peaked around age 22. Interestingly, usage patterns suggest a "social substitution effect," where users increasingly replace human support networks with AI agents, particularly when seeking mood modification or relief from social anxiety.
Conclusion and Future Work
The CAIDS-20 provides a reliable and valid tool for assessing the psychological impact of conversational AI. While the current study focused on Chinese college students, the findings highlight a growing need for cross-cultural validation. Future research should distinguish between productivity-oriented AI use and social/romantic chatbot dependence, and incorporate objective metrics such as neuroimaging to further validate the neurocognitive basis of AI addiction.
Summary updated