Patterns of Engagement With an AI Conversational Agent for Mental Health and Associations With Anxiety and Depression: A Cross-Sectional Study

At a glance

Key findings

• The study introduced pattern-based engagement metrics — interaction depth and temporal consistency — that capture how, not just how much, people use an AI conversational agent.
• Four engagement profiles emerged among 112 users who completed at least five sessions and a linked survey.
• Engagement patterns were significantly associated with self-reported anxiety (GAD-7) and depression (PHQ-8), over and above total time spent on the platform.
• Because the study is cross-sectional, the associations cannot establish cause — and more consistent use was linked to higher, not lower, reported symptoms.

What the study examined

This study asked whether how people engage with an AI conversational agent — not just how much — relates to their mental health. Survey responses were linked to backend usage data for 112 users who had completed at least five sessions.

Symptoms of depression and anxiety were measured with the PHQ-8 and GAD-7 — both validated self-report scales.

A new way to measure engagement

Most digital mental health research counts sessions or total minutes. This study instead characterized two dimensions that volume metrics miss: interaction depth (how long and sustained sessions tend to be) and temporal consistency (how evenly usage is spread across days).

Crossing these two dimensions produced four engagement profiles: extended and episodic, extended and consistent, brief and episodic, and brief and consistent.

What the researchers found

Engagement patterns were significantly associated with self-reported symptoms, even after accounting for age, gender, and total time on the platform. Users with longer, clustered sessions (extended and episodic) reported the fewest anxiety and depression symptoms; users with brief, evenly spaced sessions (brief and consistent) reported the most.

How to read these findings

This is a cross-sectional study, so it cannot show that one thing causes another. A likely explanation runs the other way: people experiencing more symptoms may reach for frequent, brief support more often. The association does not mean that using the app more — or less — changes symptom levels.

The value of the study lies in the measurement approach: pattern-based metrics reveal differences between users that simple session counts and total time cannot.

Why it matters

For a field that usually equates “more usage” with “better,” these results suggest engagement is multidimensional. Understanding the shape of engagement — not just its volume — may help future tools recognize when someone is struggling, make stronger usage recommendations, and tailor support accordingly.

The authors describe the findings as preliminary and hypothesis-generating, and call for prospective studies with larger, more diverse samples to test directionality.

Limitations

The authors describe these findings as preliminary and hypothesis-generating. A few constraints are worth keeping in mind when reading them:

• Cross-sectional design: the study measured associations at a single point in time and cannot establish cause or direction.
• Small, self-selected sample (n = 112) that was predominantly White and male, limiting generalizability.
• Survey completion was incentivized and optional, which can introduce response bias.
• The engagement metrics and the ≥5-session threshold are novel analytical choices that would benefit from independent validation.

Citation & declarations

More research