Generative Experiences for Digital Mental Health Interventions: Evidence from a Randomized Study
Ananya Bhattacharjee, Michael Liut, Matthew Jörke, Diyi Yang, Emma Brunskill
cs.HC, cs.AI
2026-04-09
GUIDE has an LLM generate both content and runtime interaction structure for stress support; in a preregistered N=237 trial it beat an LLM baseline on stress reduction and UX.
Digital mental-health (DMH) tools mostly personalize what content they show: a meditation or a reframing exercise picked to match your mood or context. But the interaction format is fixed. The same text box, the same steps, every time. This paper argues that well-matched content still fails when the format does not fit how someone can engage right now. A person who is overwhelmed may not have the energy to type a long reflection; someone commuting can only take in audio. A fixed template misses the layer where the support is actually delivered.
GUIDE's move is to make how the support is experienced something the system generates at runtime, not something hardcoded.
GUIDE is a three-stage pipeline, and each stage uses the same trick: generate several candidates, score them against a rubric, keep the best.
First, context elicitation. Five guided questions walk from "what happened" through difficulty, impact, sense of control, and current setting. Users answer by text or voice and get an editable summary.
Second, intervention selection. It generates three CBT-based candidates, some thought-focused (cognitive restructuring) and some action-focused (behavioral activation, problem solving), then scores each on eight rubrics: narrative flow, safe sequencing, alignment with psychological principles, everyday feasibility, and so on. The highest-scoring candidate wins.
Third, the part that matters: UX composition. GUIDE keeps a set of interaction primitives (text input, choice, audio message, timed guided sequence, image, timer) and assembles them into a full flow for the chosen intervention, scored against seven rubrics such as intervention-interface alignment, task efficiency, usability, and information clarity.
So the modality is itself generated. The same user might be guided through text in one session and led by voice and a timer the next. Sessions averaged 18.6 minutes, with a mean end-to-end generation latency of about 52 seconds.
A preregistered, between-participant, single-session stress-management study with 237 undergraduates (mean age 20.78; 122 GUIDE, 115 control). The control was itself an LLM cognitive-restructuring system: describe the situation, rank and identify one of 13 predefined "thinking traps" with a fine-tuned model, then write a more balanced reframed thought. So this isolates the increment from generating the experience format over generating content alone, not AI versus nothing.
| Metric | GUIDE | Control | p | Cohen's d |
| Stress reduction | 0.65±0.7 | 0.35±0.8 | .02 | 0.39 |
| User experience | 0.49±0.6 | 0.33±0.6 | .04 | 0.27 |
| Felt understood | 3.70±0.9 | 3.43±1.0 | .04 | 0.30 |
| Intent to reuse | 3.26±1.0 | 2.90±1.2 | .03 | 0.33 |
| Enjoyment | 3.44±1.0 | 3.16±1.0 | .04 | 0.28 |
| Felt personalized | 3.39±1.1 | 3.40±1.1 | .60 | -0.01 |
The primary outcomes clear 0.05, with small-to-medium effects (d=0.27 to 0.39). The interesting parts are two contrasts. First, interaction diversity: GUIDE produced 76 distinct sequences across 122 sessions (normalized entropy 0.87, average similarity 0.40 versus a fixed 1.00 for control). Second, subjective personalization was a null (p=.60). Objectively very different, subjectively not felt, a tension the authors flag themselves. In a subgroup, the 17 participants who got cognitive restructuring through GUIDE showed larger stress reduction (M=0.88±0.60, p=.002, d=0.74), but n=17 makes it a lead rather than a finding. A regression controlling for pre-intervention stress still favored GUIDE (β=-0.28, p=.006).
The "generative experience" idea is not specific to mental health. Any setting where one interface cannot serve every user's current state (onboarding, coaching, online education) can borrow it: let the model pick the modality and assemble the flow, instead of emitting a wall of text. For people building agents, it is a clean example of models orchestrating UI, built from familiar parts: best-of-N with a rubric judge, plus a small DSL of UX primitives.
The null on perceived personalization is worth holding onto too. However varied the output, users do not automatically feel more cared for. Personalization has to be validated at the perception layer, not just in the distribution.
The authors list several: single session, student sample, underpowered exploratory analyses, a multi-component design that cannot attribute the gain to any one part, only two raters for judging, and inconsistent perceived personalization across sequences.
A few more from reading it. The control is itself an LLM system, so this shows the increment of generating the experience format over generating content, not over standard care or no intervention; we do not know whether GUIDE beats established evidence-based DMH apps. The sample is computer-science undergraduates, digitally fluent and dealing with everyday stress rather than clinical disorder, so generalizing to patients is uncertain. The 52-second latency is long for a real-time interaction. And effects of d=0.39 rest on a single session of self-report, which is an immediate stress drop, not evidence of clinical value over time.