Depression Symptoms and Relational Patterns in 187k ChatGPT Histories

This study investigates how individuals with varying levels of depressive symptoms use ChatGPT as an informal support infrastructure. By analyzing over 187,000 conversation histories from 766 participants who completed the PHQ-8 (a measure of depressive symptom severity), the researchers sought to understand the topics, timing, and nature of these interactions. The study frames ChatGPT not as a clinical tool, but as a persistent, private, and always-available system that fills gaps when traditional support is inaccessible.

Participants scoring at or above the moderate-symptom threshold (PHQ-8 ≥ 10) exhibited distinct usage patterns compared to those with lower scores. They engaged in more conversations centered on mental health, loneliness, and interpersonal problems, and used more first-person singular pronouns and absolutist language. Notably, these users were more likely to interact with the system during late-night hours (23:00–04:59) and in recurring monthly patterns. While these users entered high-disclosure and support-seeking contexts more frequently, the researchers found that ChatGPT’s rate of professional redirection—such as suggesting a user contact a healthcare professional—did not increase in response to these higher-need interactions.

As LLMs become increasingly integrated into daily life, they are functioning as a form of "informal support infrastructure." This research highlights a critical design tension: while ChatGPT provides immediate, non-judgmental validation, it lacks the professional boundaries and clinical sensitivity required to manage high-disclosure mental health interactions. The findings suggest that developers should consider session-aware and history-aware features that can detect when a user is in a high-need state and adjust the system's response style accordingly, rather than relying on a one-size-fits-all approach.