JMIR Publications

ABSTRACT

Background:

Anxiety is among the most common mental health problems experienced by university students, while university counseling services face demand that human staff alone cannot meet. Chatbots based on artificial intelligence (AI) and large language models (LLMs) are increasingly offered as a solution, but evidence on how AI works alongside counselors remains scattered.

Objective:

This systematic review aims to synthesize evidence on the functions of AI in anxiety support, human-AI collaboration models, and safety and ethical requirements, and then derive a framework for university counseling services.

Methods:

Searches were conducted in Scopus, PubMed, ERIC, and Lens.org on October 28, 2026, for English-language journal articles published 2021-2026 examining AI systems for mental health support and involving university students, young adults, counselors, or counseling services. Reporting followed PRISMA 2020. Risk of bias was assessed using the JBI checklist according to design, the Mixed Methods Appraisal Tool 2018, and PROBAST. Findings were synthesized thematically and narratively using Whittemore and Knafl's integrative approach.

Results:

Of 368 records, 75 studies reported in 76 articles met the inclusion criteria: 8 qualitative studies, 11 cross-sectional and survey studies, 8 comparative experimental and observational studies, 18 system development, mixed methods, and prediction model studies, and 30 reviews and conceptual papers. Only 14 studies measured anxiety directly. Qualitative and survey studies consistently position AI as a complement and expect human judgment in crises. Experimental studies show short-term anxiety reduction; a systematic review of 31 trials reported an SMD of -0.37 (95% CI -0.58 to -0.17), with the greatest improvements in clinician-supported AI-assisted CBT. System development and prediction model studies report high performance, but mostly on simulated or public data and at high risk of bias. Risk of bias was low in 24 studies, moderate in 30, and high in 21.

Conclusions:

Evidence supports AI as a risk-tiered complement within counselor-controlled services, rather than as a standalone chatbot. We propose a Human-AI Collaborative Anxiety Support (HACAS) framework with four tiers of service, counselor decision-making at every escalation point, and a governance foundation. This framework needs prospective testing.