Facilitating Self-Guided Mental Health Interventions Through Human-Language Model Interaction: A Case Study of Cognitive Restructuring

TL;DR

This study tackles the accessibility gap in mental health care by evaluating a language-model–driven tool to guide cognitive restructuring in a real-world, large-scale field trial (N=15,531). It demonstrates meaningful reductions in emotional intensity and better overcoming negative thoughts for a majority of users, with notable variation by age, gender, and education. Through ablations and targeted adaptations (e.g., simpler language for adolescents), the work identifies design principles—contextualization, iterative interactivity, and careful safety/ethics practices—that influence efficacy and equity. The findings offer practical guidance for deploying scalable, self-guided mental health interventions that leverage AI while highlighting the need for demographic tailoring and ongoing safety considerations.

Abstract

Self-guided mental health interventions, such as "do-it-yourself" tools to learn and practice coping strategies, show great promise to improve access to mental health care. However, these interventions are often cognitively demanding and emotionally triggering, creating accessibility barriers that limit their wide-scale implementation and adoption. In this paper, we study how human-language model interaction can support self-guided mental health interventions. We take cognitive restructuring, an evidence-based therapeutic technique to overcome negative thinking, as a case study. In an IRB-approved randomized field study on a large mental health website with 15,531 participants, we design and evaluate a system that uses language models to support people through various steps of cognitive restructuring. Our findings reveal that our system positively impacts emotional intensity for 67% of participants and helps 65% overcome negative thoughts. Although adolescents report relatively worse outcomes, we find that tailored interventions that simplify language model generations improve overall effectiveness and equity.

Figures (14)

• Figure 1: We design a human-language model interaction based system for self-guided cognitive restructuring of negative thoughts. The system involves (a-b) describing the context by participants, (c-d) LM-assisted identification of thinking traps, and (e-g) LM-assisted writing of reframed thoughts.
• Figure 2: (a) Reduction in emotion intensity of participants before and after using the system (emotion scale: 1 to 10). We found that 67% of the participants reported having a positive reduction in (negative) emotions (N=1,922). (b) Participants with higher emotion intensity before using the system reported a higher reduction in emotion intensity post the system usage (N=1,922). Error bars represent 95% bootstrapped confidence intervals.
• Figure 3: Randomized controlled trial to estimate the effects of contextualizing thoughts through situation (N=1,636). (a) Contextualizing participant thoughts through their situations led to 2.80% more helpful reframes (3.31 vs. 3.22; $p = 0.0104$) but did not lead to more relatable reframes. (b) Asking for additional context did not lead to a lower completion rate. Error bars represent 95% bootstrapped confidence intervals. Effects without p-values were not significant at $\alpha = 0.05$.
• Figure 4: Randomized controlled trial to estimate the effects of enabling participants to iteratively edit reframes through increased interaction with the language model (N=2,165). 38% of participants chose to use this intervention. (a) Having the option of interactive reframe edits available to participants led to a 23.73% greater reduction in emotion intensity (2.19 vs. 1.77). (b-e) However, it did not lead to significant differences in other outcomes (at $\alpha=0.05$). Error bars represent 95% bootstrapped confidence intervals. Effects without p-values were not significant at $\alpha = 0.05$.
• Figure 5: Participants were provided with an option to seek one or more of the following types of more specific suggestions from the language model -- actionable, empathic, or personalized. Among those who used any of the options (N=992), (a) those who chose to make their reframes actionable experienced superior effectiveness across all five outcomes; (b) those who chose to make their reframes empathic reported 21.86% higher reduction in emotion intensity (2.23 vs. 1.83), 5.52% higher reframe helpfulness (3.44 vs. 3.26), and 5.14% higher skill learnability (3.48 vs. 3.31) and no significant differences based on reframe relatability and reframe memorability (at $\alpha = 0.05$); (c) those who chose to make their reframes personalized reported no significant differences in outcomes (at $\alpha = 0.05$). Error bars represent 95% bootstrapped confidence intervals. Effects without p-values were not significant at $\alpha = 0.05$.