TL;DR
This longitudinal study of 114 adults reveals that early motivations for using AI—whether for social, entertainment, or even instrumental reasons—are strong predictors of "problematic AI use" one year later. The research highlights that the drift toward using AI for emotional regulation is a far more significant risk factor for psychological dependence than simple usage frequency.
What Problem Does This Paper Address?
As conversational AI becomes integrated into daily life, people are increasingly using these systems for emotional support, self-disclosure, and companionship. While short-term evidence suggests these interactions can reduce loneliness, there is growing concern about long-term "problematic AI use"—a pattern of engagement that individuals find difficult to regulate and that interferes with their daily lives.
Most existing research is cross-sectional, providing only a snapshot of user behavior. This paper addresses the lack of longitudinal evidence, specifically investigating whether early motivations (the "why") and usage patterns (the "what for" and "how much") can serve as early warning signals for future psychological dependence.
Methods
The researchers conducted a 14-month longitudinal study involving 114 U.S. adults. The study spanned three measurement waves:
- Time 0 (T0): Pre-experiment screening and intake.
- Time 1 (T1): Post-experiment survey (2 months after T0) measuring four types of AI motivation: Instrumental (tasks), Entertainment, Social, and Escape.
- Time 2 (T2): One-year follow-up measuring Problematic AI Use (adapted from the Problematic ChatGPT Use Scale), Use Purpose (Emotional vs. Functional), and Usage Frequency.
The team employed Ordinary Least Squares (OLS) regression to predict T2 outcomes from T1 motivations, controlling for age, education, and baseline mental health. They further used mediation analysis (with 95% percentile bootstrap confidence intervals) to determine if the relationship between motivation and problematic use was driven by how much people used AI or what they used it for.
Datasets
The study relies on a primary longitudinal survey dataset of 114 participants recruited via the Prolific panel. The researchers utilized several validated psychometric scales:
- Motivation for AI Use: A 12-item scale by Huang et al. covering four subscales.
- Problematic AI Use: A 10-item adaptation of Yu et al.’s scale, focusing on features like preoccupation, loss of control, and functional impairment.
- AI Use Purpose: Items adapted from Wolf and Maier to distinguish between emotional and functional goals.
The full list of measures and their roles in the analysis can be seen in the table below:
Key Insights and Ideas
- Motivation as a Stable Trait: AI use motivation appears to be a relatively stable individual difference over time ($r \approx 0.75$), making it a viable candidate for early risk screening.
- The "Emotional Drift" Trap: The most striking finding is that the path to problematic use is primarily mediated by emotional use purpose (using AI to manage feelings or work through difficulties). This accounted for 60% of the total effect of motivation on problematic use.
- Instrumental Use is Not Inherently Protective: Even users who initially reported purely "instrumental" motivations (e.g., using AI for work or learning) were at risk if their actual behavior drifted toward emotional purposes over the year.
- Purpose > Frequency: While usage frequency does predict problematic use, it is a much weaker signal than usage purpose. Heavier usage only accounted for about 18% of the link between motivation and dependence.
- Demographics are Poor Predictors: Beyond age (where older participants reported slightly higher problematic use), individual characteristics like gender, race, education, and even pre-existing mental health conditions did not reliably predict who would develop problematic patterns.
Key Results
- Predictive Power: All four T1 motivation types significantly predicted T2 problematic use (overall $\beta = 0.367, p < .001$).
- Mediation Dominance: Emotional use purpose was the largest channel for every motivation type. For social motivation, it accounted for 77% of the total effect, compared to just 22% for usage frequency.
- Direct vs. Indirect Effects: When emotional use purpose was included in the model, the direct path from motivation to problematic use often became non-significant, suggesting that the "why" matters primarily because it leads to "emotional" usage patterns.
- Long-term Stability: The five-week experimental "nudge" to use AI for emotional purposes in the prior study had no detectable lasting effect a year later, suggesting that problematic use develops from internal motivations rather than short-term external assignments.
Limitations and Future Work
- Sample Characteristics: The sample was generally "low-risk," with few participants reaching clinical levels of problematic use. The findings may not generalize to high-severity populations.
- Sample Size: With $n=114$, the study was not sufficiently powered to detect very small effects or complex interactions between individual characteristics.
- Temporal Ordering of Mediators: Because use purpose and the outcome were measured in the same wave (T2), the study provides a descriptive decomposition rather than a strictly causal sequence for the mediation.
- Future Directions: The authors suggest that system designers should monitor shifts in use purpose (which can be inferred from logs) rather than just usage volume to identify users at risk. Future research should also investigate how specific AI design features (like personification or responsiveness) influence the drift toward emotional dependence.
Summary updated