Issue #003 — The Lancet's GBD 2023 update lands 1.2B globally with anxiety up 158% since 1990, and Spring Health's VERA-MH crystallises as the open-source safety benchmark The Path used to anchor its $14.3M launch.
Weekly Intelligence · Week 3 · 23 May 2026 · Issue #003
The Lancet's GBD 2023 update lands 1.2B globally with anxiety up 158% since 1990, and Spring Health's VERA-MH crystallises as the open-source safety benchmark The Path used to anchor its $14.3M launch.
Executive Summary
This was an epidemiology- and industry-positioning week. The most consequential signal is The Lancet's Updated trends in the global prevalence and burden of mental disorders, 1990–2023 — a Global Burden of Disease 2023 systematic analysis across 204 countries — putting the worldwide prevalence of mental disorders at ≈1.2 billion in 2023, a 95.5% increase since 1990, with anxiety and depression registering 158% and 131% growth respectively over that window. The 15-to-19 age band is, for the first time in GBD history, the peak-burden cohort. On the industry side, two interlinked announcements on 21 May reframe how mental-health AI products will go to market: Tony-Robbins-co-founded The Path exited stealth with a $14.3M seed (Prime Movers Lab–led) and positioned itself by self-reporting a score of 95 on VERA-MH, the open-source Spring Health / Expert Council evaluation that has emerged this spring as the field's first openly-licensed multi-turn clinical-safety benchmark for mental-health LLMs (distinct from last week's Mpathic clinician-built suite, which is not open-sourced). The week is short on new primary research but long on infrastructure: a population-level prevalence baseline and an evaluation primitive that challengers can now build against.
Key Metrics
| Metric | Value | Source |
|---|---|---|
| GBD 2023 global mental-disorder prevalence (2023) | 1.2 billion | The Lancet · 21 May 2026 |
| GBD 2023 anxiety / MDD growth since 1990 | +158% / +131% | The Lancet · 21 May 2026 |
| The Path seed round / claimed VERA-MH score | $14.3M / 95 of 100 | TechCrunch · 21 May 2026 |
Epidemiology
Lancet GBD 2023: global mental-disorder burden reaches 1.2B, peak shifts into the 15–19 cohort
The Institute for Health Metrics and Evaluation–led Global Burden of Disease 2023 mental- disorder update was published in The Lancet on 21 May 2026 and is the field's first major post-pandemic systematic re-baseline. The analysis covers 12 mental-disorder categories across 204 countries and territories from 1990 to 2023. Headline findings: prevalence reaches ≈1.2 billion globally in 2023, a 95.5% rise since 1990; anxiety disorders grew 158% and major depressive disorder 131% over the same period, jointly the world's two most common mental-health conditions. For the first time in GBD history the peak-burden cohort is the 15-to-19 age band, displacing the previously dominant young-adult window — a structural shift the authors flag as "a more concerning phase of worsening mental disorder burden globally." Female burden is disproportionate across the lifespan. The paper functions as the new denominator the computational-behavioral-detection literature should be sized against: the gap between detection research throughput and the at-risk adolescent population it nominally serves is now larger than any prior GBD wave has reported, and the case for scalable early-detection pipelines (passive, unobtrusive, school-deployable) — rather than clinic-bound assessment — is incrementally strengthened by the demographic shift alone.
Source: GBD 2023 Mental Disorders Collaborators · The Lancet · 21 May 2026 · 10.1016/S0140-6736(26)00519-2 Source: Brenda Goodman · CNN Health · 21 May 2026 · cnn.com
AI / ML for Mental Health Detection
VERA-MH emerges as the first open-source clinician-rubric benchmark a new entrant has scored against
The Spring Health / Expert Council benchmark Validation of Ethical and Responsible AI in Mental Health (VERA-MH) — first proposed in late 2025 and now in its 60-day Request for Comment phase — is the first openly-licensed multi-turn evaluation rubric for mental-health LLMs and is methodologically distinct from last week's Mpathic clinician-built suite (Issue #002), which is not open-sourced and is held by a single vendor. VERA-MH uses a clinically-developed rubric applied by an LLM-as-judge against synthetic multi-turn role-plays grounded in evidence-based suicide-prevention practice; preliminary validation reports an inter-clinician inter-rater reliability of 0.77 and an LLM-judge–to-clinical-consensus agreement of 0.81. Spring Health's own deployed system scored 82/100. The benchmark's structural contribution this week is not the numbers themselves but the category: a non-vendor, open-source, multi-turn safety evaluation challengers can now publicly score against — a primitive the chatbot field has so far lacked. The release timing is significant: 21 May saw the first third-party product (The Path; see Industry section) anchor a launch on a self-reported VERA-MH score, indicating the benchmark is moving from proposal to industry adoption inside one quarter.
Source: Spring Health · "Spring Health and Expert Council Release VERA-MH, the First Open-Source Evaluation for Validating AI in Mental Health" · 2026 · springhealth.com Source: VERA-MH preprint · arXiv · 2026 · arXiv:2602.05088
⚠️ Preprint — not yet peer reviewed
Industry and Product News
The Path closes $14.3M seed, exits stealth with self-reported VERA-MH score of 95
The Path, a Tony-Robbins-co-founded AI mental-health platform built by former Calm leadership (data-science / AI head Anson Whitmer and engineering head Tyler Sheaffer), exited stealth on 21 May with a $14.3M seed round led by Prime Movers Lab (where Robbins is a partner). The product offers eleven configurable AI "therapists," is free at launch with a planned $40/month tier, and has handled ≈3.5M messages across ~50,000 users since soft launch. The most relevant detail for this newsletter is positional: The Path is the first stealth-exit product this cycle to anchor its launch communication on a self-reported VERA-MH safety score (95 of 100), with the press release directly contrasting that figure against a "top score of 65 for consumer bots." The release is consequential less for the round size than for the new market entry pattern it codifies: the safety-benchmark-self-report is now a launch-narrative primitive on the same axis as efficacy claims and clinician-endorsement claims, which is the position FDA digital-health advisers and the AMA's April letter to Congress (covered as background in earlier issues) have been pushing for over the last six months. Expect competing entrants and incumbents (Wysa, Woebot, Replika, Character.AI) to publish VERA-MH scores within weeks — and a parallel pattern to emerge for the Mpathic suicide benchmark — as not doing so becomes a negative inference.
Source: Marina Temkin · TechCrunch · 21 May 2026 · techcrunch.com Source: MobiHealthNews · 21 May 2026 · mobihealthnews.com Source: HIT Consultant · 21 May 2026 · hitconsultant.net
Forward Outlook
- Near-term: Watch for Wysa, Woebot, Limbic, Character.AI, and at least one of the frontier-lab providers (Anthropic, OpenAI, Google) to publish VERA-MH scores within 4–6 weeks. The benchmark's RFC period closes during that window, and the first revised version is likely to harden score interpretability before the Mpathic vs. VERA-MH dual-benchmarking pattern settles. Expect a small wave of methodology critiques of VERA-MH's LLM-as-judge step, mirroring the analogous critique cycle that hit medical LLM evaluation in late 2025.
- Mid-term: The GBD 2023 adolescent-peak finding will be cited heavily in funding and policy arguments for school-deployed passive-sensing and digital-phenotyping pipelines (Mindcraft, mindLAMP-schools, Beiwe-schools) over the next 12 months. The IHME data is the strongest current denominator for "the cohort we are nominally trying to detect early." Expect at least one major US or UK national funder to issue an adolescent-cohort-specific digital biomarker RFP citing the GBD figures by year-end.
- Long-term: The combined GBD trajectory (+95.5% prevalence over 33 years, with pandemic-era step changes that have not reverted) and the structural deficit of clinical-workforce capacity make a triage-grade AI screening layer the most plausible high-impact deployment shape — pushing the field's normative target away from autonomous treatment and toward signal-routing into human-clinician care. VERA-MH's design (multi-turn, safety-first, open-source rubric) is well-aligned with that shape; the Mpathic safety benchmark is the parallel commercial-vendor anchor for the same evaluation surface.
Sources used: 8 · Week 3 · Next issue: 30 May 2026