Clinical promise, automated empathy and the uncertain responsibility behind digital mental health care.
Artificial intelligence is entering anxiety care without always being recognized as part of healthcare. It appears through symptom trackers, conversational agents and applications that offer immediate support in language that can feel private and attentive. This availability responds to a real deficit: anxiety disorders affected an estimated 359 million people in 2021, while approximately one in four people in need receive treatment [1]. The attraction is therefore understandable, although immediate access and clinical protection do not necessarily advance together.
Evidence offers grounds for interest, but not for equivalence with psychotherapy. A 2024 meta-analysis found a small short-term reduction in anxiety symptoms among users of AI-based conversational interventions, with no significant effect at the three-month follow-up [2]. Such findings support bounded functions—including psychoeducation, symptom monitoring and structured exercises—while leaving open whether the benefit comes from artificial intelligence itself, the intervention design, the novelty of the exchange or the surrounding clinical architecture.
That uncertainty becomes visible in the Therabot trial. The purpose-built generative system produced short-term improvements among adults with mental health symptoms, including generalized anxiety [3]. Yet Therabot was not an unrestricted consumer chatbot. It was clinically curated, monitored by researchers and supported by safety procedures; staff intervened when inappropriate responses or risk signals appeared. Its effectiveness cannot be separated easily from those human arrangements. What looked like autonomous treatment remained, in part, supervised care delivered through an artificial interface.
To the user, however, the distinction may be difficult to perceive. Clinical and general-purpose systems can both produce coherent, personalized and emotionally responsive language. Experimental evidence indicates that revealing whether personal narratives are human- or AI-generated changes users’ empathic responses [4]. In mental health interactions, transparency may therefore alter the emotional meaning of the conversation. Linguistic warmth can encourage disclosure, but it can also create an impression of understanding that exceeds the system’s actual competence and responsibilities.
Anxiety complicates this relationship because reassurance is not always therapeutically neutral. It can reduce distress briefly while reinforcing the avoidance pattern that sustains it. A person who repeatedly asks a chatbot to interpret physical sensations, intrusive thoughts or social encounters may experience temporary relief without developing tolerance for uncertainty. General-purpose chatbots may facilitate reassurance-seeking and dependence in ways that perpetuate anxiety and obsessive-compulsive processes [5]. No single response must be obviously harmful for the interaction to become clinically counterproductive.
More visible failures appear when a system does not recognize that the conversation has crossed into crisis. Panic symptoms can overlap with medical emergencies, and anxiety may coexist with depression, trauma, substance use or suicidal risk. An evaluation of 29 mental health chatbot agents found significant deficiencies in responses to escalating suicidal ideation, including weak contextual interpretation and incomplete emergency guidance [6]. The troubling feature is not simply that a chatbot may respond incorrectly. It may continue responding persuasively after a clinician would have changed the nature of the intervention.
Responsibility becomes diffuse at that point. Developers may characterize the system as informational, platforms may rely on disclaimers, professionals may recommend tools they cannot continuously evaluate, and users may disclose intimate information without fully understanding how it will be processed. The World Health Organization emphasizes transparency, risk assessment, accountability and meaningful human control in the use of advanced AI systems for health [7]. These principles are substantial, but they do not by themselves identify who acts when an apparently supportive exchange begins to produce harm.
Protection cannot be reduced to an emergency notice displayed beneath the conversation. It depends on defined clinical boundaries, effective escalation mechanisms, privacy safeguards, adverse-event monitoring and validation tied to a particular system, version, population and purpose [8]. Even these conditions remain unstable because updates to memory, prompts, safety layers or commercial engagement strategies may alter system behavior after evaluation. Human oversight can reduce this uncertainty, although it may also become ceremonial when no identifiable person has the authority or time to intervene.
Completely excluding artificial intelligence would preserve many of the barriers that already limit mental health care. Carefully delimited systems may offer useful low-intensity support to people who otherwise receive none. Wider availability, however, can coexist with unequal access, relational dependence and fragmented accountability [8]. The question is no longer simply whether AI can generate therapeutic language. It is whether someone remains answerable when that language begins to shape behavior, delay professional assistance or acquire the meaning of care in the mind of the patient.
References
[1] World Health Organization. (2025, September 8). Anxiety disorders. World Health Organization.
[2] Zhong, W., Luo, J., & Zhang, H. (2024). The therapeutic effectiveness of artificial intelligence-based chatbots in alleviation of depressive and anxiety symptoms in short-course treatments: A systematic review and meta-analysis. Journal of Affective Disorders, 356, 459–469. doi:10.1016/j.jad.2024.04.057
[3] Heinz, M. V., Mackin, D. M., Trudeau, B. M., Bhattacharya, S., Wang, Y., Banta, H. A., Jewett, A. D., Salzhauer, A. J., Griffin, T. Z., & Jacobson, N. C. (2025). Randomized trial of a generative AI chatbot for mental health treatment. NEJM AI, 2(4), Article AIoa2400802. doi:10.1056/AIoa2400802
[4] Shen, J., DiPaola, D., Ali, S., Sap, M., Park, H. W., & Breazeal, C. (2024). Empathy toward artificial intelligence versus human experiences and the role of transparency in mental health and social support chatbot design: Comparative study. JMIR Mental Health, 11, Article e62679. doi:10.2196/62679
[5] Golden, A., & Aboujaoude, E. (2026). A transdiagnostic model for how general-purpose AI chatbots can perpetuate OCD and anxiety disorders. npj Digital Medicine, 9, Article 343. doi:10.1038/s41746-026-02531-7
[6] Pichowicz, W., Kotas, M., & Piotrowski, P. (2025). Performance of mental health chatbot agents in detecting and managing suicidal ideation. Scientific Reports, 15, Article 31652. doi:10.1038/s41598-025-17242-4
[7] World Health Organization. (2024). Ethics and governance of artificial intelligence for health: Guidance on large multi-modal models. World Health Organization.
[8] López Ayala, M. G. (2026). Artificial intelligence and anxiety care: Clinical promise and ethical boundaries. Revista Lince de Ciencias Sociales, Humanidades y Tecnologías, 2(1), 111–123. doi:10.63622/RLI/2026.01/06