Real-World Use of a Mental Health AI Companion: Multiple Methods Study
Bibliographic record
Abstract
BACKGROUND: The rapid acceleration of large language models (LLMs) creates opportunities to expand the accessibility of mental health support; however, general artificial intelligence (AI) tools lack safety guardrails, evidence-based practices, and medical regulation compliance, which may result in misinformation and failing to escalate care in crises. In contrast, Ebb, Headspace's conversational AI tool (CAI tool), was purpose-built by clinical psychologists and research experts using motivational interviewing techniques for subclinical guidance, incorporating clinically backed safety mechanisms. OBJECTIVE: This study aimed to (1) understand Headspace members' sentiment toward AI and expectations for a mental health CAI tool, (2) evaluate real-world use of Headspace's CAI tool, and (3) understand how members perceive a CAI tool fitting into their mental health journey. METHODS: This was a multiple method study using three data sources including Headspace members: (1) cross-sectional survey (n=482) assessing demographics, AI use, and the Artificial Intelligence Attitude Scale-4 (AIAS-4); (2) real-world engagement descriptive analysis (n=393,969) assessing session and message counts, retention, and conversation themes; and (3) diary study (n=15) exploring the CAI tool's role within members' mental health journey. App engagement was compared between CAI tool 1.0 and CAI tool 2.0, where CAI tool 2.0 featured enhanced LLM conversational prompts, comprehensive memory, woven content recommendations, and more robust safety detection. RESULTS: While the majority of survey respondents used and would continue to use general AI tools, overall attitudes toward AI remained neutral (AIAS-4 mean 5.7, SD 2.2, range 1-10). Survey results suggest that members viewed the CAI tool as a guide to navigate to mental health resources and Headspace content and provide in-the-moment support. Members emphasized the need for data safety and ethics transparency, clinical guidelines structure, and for the CAI tool to be a resource in addition to human-delivered mental health care, not a replacement. Real-world CAI tool use showed strong engagement across 393,969 Headspace members. The product evolution to CAI tool 2.0 led to increased retention (77,894/153,249, 50.8% completed 2 sessions within 7 days vs 68,701/240,720, 28.5% for CAI tool 1.0) and higher positive conversation ratings (37,819/40,449, 93.5% vs 94,308/104,323, 90.4%). Retained CAI tool 2.0 users showed greater retention (6.1 sessions per user) compared to all CAI tool 2.0 users (2.9 sessions per user) and CAI tool 1.0 (2.4 sessions per user). Diary study results suggest that members imagined using the CAI tool when feeling stress or anxiety and during morning routines, commutes, or while winding down at night. CONCLUSIONS: Results emphasize the necessity of research-backed, purpose-built mental health AI products with minimum viable safeguards, including (1) transparent labeling of intended use, benefits, and limitations; (2) safety by design principles to monitor for overuse, detect risk, and flag needs for escalation; and (3) child and adolescent safeguards.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
How this classification was reachedexpand
Full frame machine prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. The Gemma side is a direct model label for every work in the frame, read from the title-only record. The Codex side is a classifier learned from the 10,348 direct Codex labels and calibrated to design-weighted sample rates; fields without enough sample support carry no Codex call. Candidate is the union of the two sides; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels.
Distilled classifier scores by category (both heads)
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.010 | 0.022 |
| Meta-epidemiology (narrow) | 0.000 | 0.001 |
| Meta-epidemiology (broad) | 0.001 | 0.001 |
| Bibliometrics | 0.002 | 0.001 |
| Science and technology studies | 0.002 | 0.001 |
| Scholarly communication | 0.002 | 0.002 |
| Open science | 0.001 | 0.002 |
| Research integrity | 0.001 | 0.001 |
| Insufficient payload (model declined to judge) | 0.004 | 0.001 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from itClassification
machine, unvalidatedMachine predicted; a candidate call from one source (direct Gemma or distilled Codex), not a consensus.
How this classification was reached, model by model and score by score, is at the end of the page under "How this classification was reached".