🤖 AI

Depression Symptoms and Relational Patterns in 187k ChatGPT Histories

This study analyzes 187,000 ChatGPT conversations to reveal that individuals with depressive symptoms use the AI more frequently for mental health and support-seeking discussions, particularly during late nights, though language patterns alone are insufficient for clinical screening, suggesting LLMs function primarily as an emerging informal support infrastructure rather than a diagnostic tool.

Original authors: Neil K. R. Sehgal, Dunigan Folk, Lyle Ungar, Sharath Chandra Guntuku

Published 2026-07-08

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Neil K. R. Sehgal, Dunigan Folk, Lyle Ungar, Sharath Chandra Guntuku

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you have a 24/7 digital diary that never sleeps, never judges, and is always ready to chat. That's what ChatGPT has become for many people. But what happens when someone is feeling down, lonely, or struggling with depression? Do they treat this digital diary differently than someone who is feeling fine?

A team of researchers from the University of Pennsylvania decided to find out. They looked at the private chat logs of 766 people who had also taken a standard questionnaire about their mood (called the PHQ-8). They compared the chats of people with higher depression scores against those with lower scores.

Here is what they found, broken down into simple concepts:

1. The "Late-Night Confessional" Effect

Think of the chat history as a nightlight.

What they found: People with higher depression scores were much more likely to be chatting with the AI late at night (between 11:00 PM and 5:00 AM).
The Analogy: While most people are asleep, these users were awake, talking to the AI. It's like a digital version of those 3:00 AM thoughts that keep you awake, except instead of staring at the ceiling, they were typing them out. They also tended to return to these heavy topics month after month, like a recurring guest at a party who always shows up to talk about the same sad story.

2. The "Heavy Backpack" of Topics

If you imagine a conversation as a backpack, people with higher depression scores were carrying a much heavier one.

What they found: Their conversations were more likely to be about loneliness, relationship problems, self-blame, and asking for emotional support. They used more "I" words (like "I feel," "I am") and absolute words (like "always," "never," "nothing").
The Analogy: Instead of asking the AI, "How do I fix a leaky faucet?" (a neutral task), they were asking, "Why does no one love me?" or "I can't do anything right." They were using the AI as a safe harbor to unload their emotional baggage when human friends or doctors weren't available.

3. The "Polite but Silent" AI

This is the most critical part of the study. The researchers wanted to know: Does the AI know when to call a professional?

What they found: Even though people with higher depression scores were sharing much more personal, painful, and "high-risk" information, the AI did not significantly change its behavior to suggest they see a doctor.
The Analogy: Imagine a very polite butler. If you tell the butler, "I'm feeling terrible," and then later, "I'm feeling worse," and then, "I'm in pain," the butler keeps nodding, saying, "I understand," and offering comfort. But the butler never picks up the phone to call a doctor, even though the situation seems to be getting more serious. The AI was warm and supportive, but it didn't "escalate" the conversation to a professional, even when the user's distress was obvious.

4. The "Crystal Ball" That's Too Fuzzy

The researchers tried to build a computer program that could look at a chat log and guess, "This person is depressed."

What they found: The computer could guess slightly better than flipping a coin, but it wasn't good enough to be useful.
The Analogy: It's like trying to diagnose a broken leg by looking at a blurry photo of a shoe. You can see something is wrong, but you can't be sure enough to send the person to the hospital. The researchers concluded that we cannot use these chat logs to screen for depression in a clinical setting. The signal is too weak and the noise is too loud.

The Big Takeaway

The paper argues that we shouldn't treat these chat logs as medical records. Instead, we should see them as evidence that people are using AI as a new kind of "informal support system."

It's a place where people go when they are lonely, when it's 3:00 AM, and when they need to talk to someone. The AI is becoming a digital friend that is always there, but it's not a doctor. The study suggests that while the AI is great at listening, it needs better design to know when a user is in crisis and needs to be gently guided toward a real human professional, rather than just offering more comfort.

In short: People with depression use the AI more at night to talk about their pain, but the AI stays polite and supportive without realizing it needs to call for backup. And while the AI's language changes slightly based on the user's mood, it's not smart enough to act as a medical detector.

Technical Summary: Depression Symptoms and Relational Patterns in 187k ChatGPT Histories

Problem Statement
Large Language Models (LLMs) like ChatGPT are increasingly utilized as private, always-available conversational systems. While prior research in CSCW and HCI has examined how individuals use networked systems for disclosure and peer support, little is known about how people with depressive symptoms interact with non-human, always-on conversational agents. Unlike traditional online support communities, LLMs offer one-on-one interaction without peers, moderators, or clinicians. This study addresses the gap in understanding how individuals with varying levels of depressive symptom severity utilize ChatGPT, specifically investigating whether these interactions function as an emerging "informal support infrastructure" and whether LLM responses appropriately adapt to user distress, particularly regarding professional redirection.

Methodology
The study analyzes a dataset of 187,093 ChatGPT conversations from 766 participants recruited via Prolific (US, UK, Canada) in Spring 2026.

Symptom Measurement: Participants completed the Patient Health Questionnaire-8 (PHQ-8). The study compares two groups: those with scores < 10 (below moderate symptom threshold, $N=571$ ) and those with scores ≥ 10 (moderate-or-greater symptoms, $N=195$ ). The PHQ-8 is treated as a measure of recent symptom severity, not a clinical diagnosis.
Data Processing: The analysis combines deterministic metrics (local-time usage, conversation length) with lexical features and LLM-derived annotations.
- LLM Annotations: Using gpt-4o-mini (temperature 0), the authors generated labels for topics (e.g., mental health, interpersonal problems, loneliness), help-seeking types, disclosure levels, support-seeking, professional redirection, and sycophancy-style responses (uncritical agreement, obsequiousness, excitement).
- Scope: Topic and conversation-level labels were applied to the full corpus. Turn-level and response-side labels were applied to targeted subsets (Health/Mental Health conversations and random non-health samples).
Statistical Analysis:
- Comparisons: Participant-weighted Welch tests, Cohen's $d$ , and Benjamini-Hochberg FDR-adjusted $q$ -values were used to compare groups.
- Sensitivity Models: Conversation-weighted models adjusted for demographic, usage, topic, model-family, calendar-time, and history-position covariates.
- Language Analysis: Differential Language Analysis (DLA) utilized 1-grams, LIWC2022, and LDA features across user, assistant, and combined text.
- Prediction: Regularized logistic regression with repeated stratified 5-fold cross-validation was used to test if language alone could predict PHQ ≥ 10 status.

Key Findings

Content and Disclosure: Participants with higher PHQ scores ( $\ge$ $\geq$ 10) brought significantly more mental-health, interpersonal, loneliness, and negative self-focused concerns to ChatGPT.
- Mental Health Share: 5.7% vs. 3.1% for the lower PHQ group.
- Disclosure: Higher-PHQ participants exhibited higher disclosure levels (1.71 vs. 1.48) and a higher share of support-seeking conversations (11.3% vs. 5.3%).
- Linguistic Markers: Higher-PHQ users employed more first-person singular pronouns (30.83 vs. 25.32 per 1k words) and absolutist terms (4.33 vs. 3.68 per 1k words).
Temporal Patterns: Higher-PHQ participants showed distinct usage timing:
- Nocturnal Use: A larger share of conversations occurred between 23:00 and 04:59 (14.2% vs. 9.7%).
- Recurring Patterns: Among active users, higher-PHQ participants had more months where mental-health conversations exceeded 5% of their monthly usage.
Response and Redirection:
- Professional Redirection: Despite higher disclosure and support-seeking, professional redirection rates were not significantly different between groups (16.9% for PHQ ≥ 10 vs. 16.0% for PHQ < 10).
- Sycophancy: Differences in sycophancy (uncritical agreement, obsequiousness) were weak and not robust after adjusting for covariates. The data suggests ChatGPT does not generally treat higher-PHQ users more sycophantically; rather, these users simply enter contexts where validation and boundaries are more consequential.
Predictive Capability: Language-based prediction of PHQ status was modest. The best model (using user LIWC features) achieved an AUROC of 0.591, with precision of 0.341 and recall of 0.487 at the default threshold. The authors conclude this is insufficient for clinical screening or triage.

Contributions

Empirical Analysis: The study provides the first large-scale linkage between survey-measured depressive symptoms (PHQ-8) and longitudinal LLM conversation histories (187k+ conversations).
Conceptual Framing: It frames ChatGPT not merely as a productivity tool but as an emerging informal support infrastructure that is private, persistent, and available outside ordinary hours, distinct from peer communities or clinical settings.
Methodological Caution: The paper demonstrates that while language signals of distress are detectable, they are too weak to justify clinical screening based on private LLM histories. It also highlights the necessity of analyzing both user and assistant turns together, as assistant responses co-vary with user symptom severity.

Significance and Claims
The paper argues that LLMs are increasingly functioning as informal support infrastructures, particularly for individuals experiencing depressive symptoms who may lack access to human support or find disclosure risky. The findings reveal a "boundary setting gap": while users with higher symptoms engage in more high-disclosure and recurring support-seeking interactions, the system's professional redirection does not scale robustly with this need.

The authors emphasize that these histories should not be treated as clinical screening data due to the modest predictive power of language models. Instead, the data serves as evidence that LLMs are filling a relational void. The study suggests design implications for recognizing high-disclosure or recurring support-seeking patterns to provide warm, bounded responses without overstepping professional boundaries or offering unverified clinical advice. The authors explicitly state that their observational data cannot determine if LLM use substitutes for or supplements human support, nor can it infer causality regarding symptom changes.