Feasibility of human-in-the-loop multimodal generative artificial intelligence chatbot for school-based adolescent mental health support

Key Findings

• — Of 1,287 first-year junior-high students screened in May 2025, 62 with elevated depressive symptoms were pragmatically assigned to the chatbot (30 initiated) or a waitlist (32); 26 of 30 (86.7%) completed all scheduled sessions with a mean total use of 101.3 minutes.
• — Mean Working Alliance Inventory-Short Revised score was 3.56 (SD 0.86).
• — Nine of the 30 intervention students triggered dangerous-behaviour alerts during the two weeks and were followed up offline by school counsellors under the protocol's alert-review and escalation pathway.
• — No clinical outcome differed between groups after baseline adjustment and correction for multiple comparisons.
• — Role separation was implemented so that company staff did not have access to identifiable interaction data, monitoring dashboards or alerts; the platform used a single binary alert criterion in this deployment.

Methodology Notes

Pragmatic (non-random) allocation based on school scheduling, small groups (30 versus 32), baseline severity imbalance, multiple outcomes, brief follow-up; co-authors include staff of the platform companies (SituTech AI Lab and Beijing Ruihong Embodied Intelligence Robot Technology), with a stated role-separation and conflict-of-interest structure. Received 2 June 2026; accepted and posted as an in-press article on 8 September 2026 (DOI 10.5498/wjp.123904). The wjgnet.com landing page returns HTTP 403 to automated fetchers; the in-press record and PDF were verified through the publisher's F6Publishing system.

Sources