JMIR Publications

ABSTRACT

Background:

Cognitive reappraisal is a widely studied emotion regulation strategy that helps individuals reinterpret stressful situations in ways that reduce their emotional impact. Digital mental health (DMH) tools often struggle to support this process because scripted templates fail to adapt to the varied and incomplete ways users describe their stressors. Large language models (LLMs) offer opportunities to increase conversational flexibility while preserving structured intervention steps.

Objective:

This study examined the feasibility of an LLM-based single-session intervention for workplace stress reappraisal. We aimed to assess whether the activity would be associated with short-term improvements in stress-related outcomes, and what design tensions arise during user interaction.

Methods:

We conducted a feasibility study with 100 employees from a large US technology company. Participants completed a structured cognitive reappraisal session delivered by a GPT-4o–based chatbot within Qualtrics. Pre–post measures included perceived stress intensity, stress mindset, perceived demand, and perceived resources (all 5-point scales). Paired Wilcoxon signed-rank tests were used with Benjamini-Hochberg correction. To complement self-reports, we analyzed sentiment and stress trajectories across conversation quartiles using a RoBERTa sentiment classifier, a RoBERTa stress classifier, and an LLM-based stress rater. Open-ended responses were analyzed using thematic analysis.

Results:

Participants wrote an average of 12.81 ± 1.66 messages, contributed 283.74 ± 243.16 words, and spent 23.09 ± 23.99 minutes engaging with the chatbot. Significant reductions were observed in perceived stress intensity (Δ = 0.29 ± 0.83, p = 0.002, r_rb = 0.54) and significant improvements in stress mindset (Δ = 1.70 ± 4.37, p = 0.002, r_rb = 0.44). Perceived resources increased (Δ = 0.17 ± 0.83, p = 0.07, r_rb = 0.32), and perceived demand decreased (Δ = 0.12 ± 0.83, p = 0.17, r_rb = 0.22) though neither reached significance. Sentiment and stress classifiers showed consistent declines in negative sentiment and stress from conversation start to end (all omnibus Friedman tests p < 0.001; Q1 to Q3 differences significant across all models). Qualitative analysis showed that participants valued the structured prompts for organizing thoughts, gaining perspective, and feeling validated. Reported design tensions included perceived scriptedness, variable preferences for conversation length, and mixed reactions to AI-driven empathy.

Conclusions:

An LLM-enhanced cognitive reappraisal activity showed promise to be delivered as a brief digital intervention and is associated with short-term improvements in perceived stress and stress mindset. Participants appreciated the clarity and reflection supported by the structured sequence, while noting important design challenges in balancing structure with conversational naturalness and contextual depth. These findings highlight both the promise and the design constraints of integrating LLMs into DMH interventions for workplace settings.