Are AI mental health companions ready for clinical or public health use?

Can AI companions actually engage users and expand access to mental health support?

Yes, the evidence shows that well-designed AI companions can attract and retain users at scale, which is a meaningful step toward addressing the global shortage of mental health professionals. In a real-world study of Headspace's purpose-built AI tool (Ebb), over 393,000 members used the tool, and the updated version (2.0) doubled the proportion of users who returned for a second session within a week—from 28.5% to 50.8% [2]. That means half of new users came back, which is strong for a digital tool. Users also rated conversations positively 93.5% of the time [2]. In a separate diagnostic study, an AI assistant called TalkToAlba conducted clinical interviews that actually outperformed standard rating scales in identifying self-reported clinician-diagnosed disorders, and most participants rated the AI as highly empathic and supportive [5]. These findings suggest that AI can feel helpful and engaging to users, and that it can be a scalable, low-cost way to reach people who might otherwise not seek care.

What are the main risks that make AI companions not ready for clinical use?

The core problem is that current AI systems lack the clinical judgment, crisis management, and accountability required for safe mental health care. A scoping review of 101 articles on conversational AI in mental health found that safety and harm was the most common ethical concern, discussed in over half (51.5%) of the papers, with top issues being suicidality, harmful suggestions, and risk of user dependency [1]. Privacy and confidentiality were flagged in 61.4% of articles, and effectiveness was questioned in 37.6% [1]. These are not minor issues—they are fundamental to whether a tool can be trusted in a crisis. A direct test of ChatGPT with three simulated patients showed that while it gave reasonable advice for a simple case, its recommendations became 'inappropriate, even dangerous' as the clinical complexity increased, because it could not ask follow-up questions or apply critical thinking [3]. Another analysis warned that ChatGPT lacks real-time fact-checking and can produce misleading information that may influence a person's thinking, potentially worsening mental health [4]. Together, these studies converge on a clear warning: AI companions can engage users, but they are not yet safe or reliable enough to replace or operate alongside human clinicians without rigorous safeguards.

What would need to change for AI companions to be ready for clinical or public health use?

The research points to three essential requirements: safety-by-design, transparent labeling, and human oversight. The Headspace study demonstrated that a purpose-built tool with clinically backed safety mechanisms—monitoring for overuse, detecting risk, and flagging needs for escalation—can achieve high engagement while maintaining a safety net [2]. The authors explicitly recommend transparent labeling of intended use, benefits, and limitations, plus child and adolescent safeguards [2]. The ethical review adds that accountability must be clearly assigned, and that AI should complement, not replace, human therapists [1]. The diagnostic study suggests that AI can be a valuable complement to traditional methods, but only when used as a screening or support tool, not as a standalone clinician [5]. In short, the path to readiness requires: (1) rigorous validation in real clinical settings, (2) built-in crisis detection and escalation protocols, (3) clear communication to users about what the AI can and cannot do, and (4) continued human oversight. None of the studies here suggest that current AI meets all these criteria.

About These Sources

This answer is built on 5 peer-reviewed studies — published from 2024 to 2026, 5 from 2024 or later, 2 in Q1 journals, collectively cited 185 times — selected as the most relevant from 5 studies that passed quality screening, drawn from 49 papers retrieved from a database of over 500 million.

Sources used in this answer

Related research questions