The results arrive as developers and health-care researchers explore whether AI could help address shortages of psychiatric professionals. Depression is widespread worldwide, yet many people cannot obtain timely psychological or psychiatric care. An AI system that can communicate continuously, screen symptoms or provide structured support could potentially extend access beyond clinics and office hours. But mental-health conversations involve highly sensitive information, including mood, trauma, medication, relationships and suicidal thoughts. People may hesitate to disclose such information to software because they do not know how it will be stored, who might access it or whether the system can be trusted to respond appropriately.

The researchers—Jingjing Tong of the University of Science and Technology of China and Wenjing Chen and David Jingjun Xu of City University of Hong Kong—examined whether anthropomorphic design could overcome some of those barriers. Anthropomorphism is the attribution of human characteristics to a nonhuman entity. In an AI health-care system, it can be introduced in several ways. Appearance anthropomorphism involves visual features, such as a human face, body, name or avatar. Emotion anthropomorphism concerns the system’s apparent emotional capacity: its use of empathetic wording, expressions of concern, reassurance, encouragement or other signals associated with human feeling. Although these two forms are often treated as part of the same design strategy, the experiment tested them separately.

The study used a 2 × 2 experimental design, meaning participants were exposed to one of four combinations: an AI doctor with low or high appearance anthropomorphism and low or high emotion anthropomorphism. This structure allowed the researchers to estimate the independent effect of each design feature as well as whether the two features influenced one another. Participants evaluated the AI doctor and reported their level of trust, their concerns about privacy and their intention to adopt or use the system. The researchers also measured perceived public stigma—the extent to which participants believed that seeking help for depression might attract negative judgment from other people.

Emotion anthropomorphism produced the clearest effect. Participants who encountered an AI doctor designed to communicate in a more emotionally humanlike manner reported greater trust and fewer privacy concerns than those who encountered a less emotionally expressive system. Trust was not merely a favorable opinion: in the study’s model, it acted as a mechanism linking design to adoption. People who trusted the AI more were more likely to say they would use it. Privacy concerns worked in the opposite direction. The more participants worried about the handling of their personal information, the less willing they were to adopt the system.

The absence of a significant effect from appearance anthropomorphism is important because visual realism is one of the most visible features in consumer AI. A lifelike avatar may attract attention, but attention is not the same as confidence in a mental-health service. A realistic face could also create expectations that the system possesses understanding or clinical judgment that it does not actually have. If its behavior fails to match its appearance, users may feel misled or experience discomfort. The study therefore points toward a distinction between looking human and behaving in a socially responsive way. For people discussing depression, conversational warmth may matter more than a polished digital face.

The researchers also identified a negative interaction between appearance and emotion anthropomorphism: when one form was high, the effect of increasing the other became weaker. In practical terms, combining a highly humanlike appearance with highly emotional communication did not simply produce an additive benefit. This may reflect a saturation effect, in which users receive diminishing social signals once an AI already appears sufficiently humanlike. It may also indicate that mismatched cues create tension. A system that looks like a person but communicates mechanically could seem uncanny, while a system that speaks warmly without pretending to possess a human body may appear more coherent and credible.

Perceived public stigma changed the strength of the emotional effect. Among participants who believed that depression carried greater social stigma, emotional anthropomorphism had a stronger influence on trust and privacy concerns. This finding is consistent with the possibility that people who fear judgment from relatives, friends, colleagues or communities may be particularly responsive to a system that appears emotionally safe. An AI may be perceived as less socially threatening than a human professional because it does not belong to the user’s immediate social environment. Yet that apparent privacy advantage is psychological, not a guarantee of technical confidentiality. An AI system can still collect, transmit or retain intimate data, and its humanlike tone should never be treated as evidence that its data practices are secure.

The study is grounded in media equation theory, which proposes that people often respond to computers and other media as though they were social actors, even when they know those systems are not human. A polite voice can elicit politeness; a conversational agent that appears attentive can invite disclosure; an expression of empathy can influence judgments about care. These responses do not require users to believe that a machine is conscious. Instead, humans automatically apply familiar social rules to cues such as language, timing and emotional expression. For mental-health AI, that psychological tendency creates both an opportunity and a hazard: carefully designed communication may make support easier to approach, but exaggerated humanlike behavior could encourage users to overestimate the system’s competence, awareness or moral responsibility.

The findings do not show that an emotionally expressive AI can treat depression, replace a psychiatrist or improve clinical outcomes. The experiment measured trust, privacy concerns and intention to adopt, not symptom reduction, diagnostic accuracy, treatment adherence or protection from self-harm. Participants’ stated willingness to use a system may also differ from behavior during a prolonged period of distress. Depression can affect concentration, motivation, decision-making and help-seeking, and the needs of a person in crisis cannot be inferred from a short interaction or a design preference. Any clinical deployment would require rigorous evaluation, transparent disclosure that the user is communicating with AI, strong privacy safeguards, pathways to qualified human care and reliable procedures for responding to emergencies.

Even with those limitations, the study offers a practical design signal in a rapidly expanding field. Developers seeking acceptance may gain more by improving emotional responsiveness, clarity and respectful communication than by investing primarily in increasingly realistic avatars. The researchers argue that AI doctors should be designed around appropriate social presence rather than superficial imitation. That means an AI could acknowledge distress without claiming to feel it, explain what it can and cannot do, provide understandable reasons for its suggestions and make it easy for users to reach human professionals. In depression care, the most persuasive illusion may not be that a machine is a person, but that the machine is attentive, accountable and safe enough to begin a difficult conversation.

While early iterations of mental health chatbots and avatars have primarily focused on demonstrating feasibility and user acceptability, the discipline now demands a pivot towards more critical, evidence-driven research. There is a growing need to unravel the clinical relevance of these digital agents beyond surface-level engagement metrics. For instance, how do therapy bots compare against active control conditions such as digital placebos or attention controls in real-world therapeutic outcomes? Understanding the nuances that distinguish genuine therapeutic change from placebo effects is vital to establishing clinical credibility and guideline integration.

Engagement is another pivotal factor warranting in-depth exploration. Unlike traditional face-to-face therapy, where sustained interaction is benchmarked by session attendance and therapeutic alliance, AI companions challenge conventional definitions of engagement. Researchers are urged to propose robust methodologies to operationalize and quantify meaningful and sustained engagement with therapy bots. Metrics that capture qualitative dimensions like emotional attunement, responsiveness, and adaptive feedback loops may offer richer insights than mere frequency or session length.

The ethical landscape surrounding AI-powered mental health tools introduces complex dilemmas. There is a pressing need to systematically examine potential risks such as data privacy breaches, inadvertent harm from inaccurate advice, and the psychological impact of human–machine boundaries becoming blurred. In particular, AI companions deployed across diverse cultural or socioeconomic contexts may inadvertently propagate biases or reinforce stigmas, necessitating vigilant scrutiny. As these digital entities increasingly influence vulnerable populations, ethical frameworks must evolve to address questions of autonomy, informed consent, and accountability in an automated care environment.

From a clinical perspective, delineating the therapeutic mechanisms encapsulated by AI companions remains a frontier topic. Are these tools simply delivering evidence-based psychotherapeutic techniques in algorithmic form, or do they uniquely modulate cognitive and affective processes through their interactive design? Emerging theories propose that AI agents may enhance self-efficacy, promote behavioral activation, or even replicate aspects of the therapeutic alliance through advanced natural language processing and empathetic response generation. Validating these hypotheses through rigorous experimental designs will provide foundational knowledge to optimize AI intervention architecture.

Equally critical is the performance evaluation of therapy bots across a spectrum of psychiatric diagnoses and demographic groups. Mental disorders vary widely in symptomatology, severity, and treatment responsiveness, which may influence how users engage with and benefit from AI-based interventions. Factors such as age, cultural background, digital literacy, and co-morbid conditions introduce layers of complexity in assessing effectiveness. Research that stratifies outcomes and tailors intervention components to diverse populations will enhance the inclusivity and generalizability of AI mental health solutions.

The integration of AI companions within broader mental health care models raises important questions about their complementary role alongside human clinicians. While some envision these tools as adjuncts augmenting traditional therapy, others propose fully automated pathways for mild to moderate conditions. Understanding optimal points for intervention, escalation protocols, and hybrid care delivery systems will be crucial for seamless incorporation into health care infrastructures. Such system-level research demands interdisciplinary collaboration including clinical psychology, computer science, health services research, and ethics.

JMIR Mental Health’s theme issue invites contributions employing cutting-edge methodologies, from randomized controlled trials to mixed-methods studies, modeling techniques, and qualitative explorations. The journal emphasizes peer-reviewed, high-impact submissions that scrutinize the complex interplay between technological innovation and mental health outcomes. Researchers are encouraged to address pressing issues such as measurement standardization, reliability and validity of AI-delivered interventions, and longitudinal impact assessments.

Given the rapid evolution of AI capabilities—embodied by advancements in machine learning algorithms, natural language understanding, and affective computing—this theme issue seeks forward-looking perspectives that anticipate future trajectories. Investigations into adaptive learning algorithms that personalize therapy content or harness real-time physiological monitoring present exciting frontiers. Moreover, assessing the scalability and sustainability of AI mental health tools in low-resource settings will have profound implications for global mental health equity.

In addition to empirical research, discourse on regulatory frameworks, reimbursement policies, and public health implications is welcomed. As jurisdictions grapple with overseeing digital therapeutic products, scholars must engage with policy dimensions to ensure safe, effective, and ethical deployment. The convergence of technological innovation and regulatory oversight ultimately shapes the societal impact of AI-driven mental health care.

Contributors interested in submitting manuscripts for this special issue should note the deadline is October 31, 2025. All papers will undergo a stringent peer-review process consistent with JMIR Publications’ standards of scientific rigor and ethical integrity. Accepted articles will be collectively published as a themed compilation titled “AI-Powered Therapy Bots and Virtual Companions in Digital Mental Health,” aiming to become a definitive reference in the field.

This initiative by JMIR Mental Health underscores the transformative potential and challenges inherent in harnessing artificial intelligence for mental health improvement. Through fostering critical and multidisciplinary scholarship, the special issue aspires to propel the responsible integration of AI companions in clinical practice, research, and policy.

Subject of Research: AI-powered mental health tools including therapy bots and virtual companions, with a focus on their effectiveness, engagement, ethical implications, and clinical integration.

News Publication Date: May 13, 2025

Web References:
https://www.jmir.org/
https://mental.jmir.org/
https://mental.jmir.org/announcements/567

Image Credits: JMIR Publications. Source: BlackJack3D

Keywords: Mental health, Artificial intelligence, Psychiatry, Cognitive robotics

The phenomenon of AI responding to emotional content stems from underlying biases, both societal and algorithmic. Just as human beings experience shifts in cognitive processing and emotional response during distressing situations, evidence suggests that large language models mirror these reactions. Studies show that negative emotional narratives exacerbate pre-existing biases within the AI systems, leading to an increase in the manifestation of potentially harmful behaviors, such as racism or sexism. Given that AI chatbots are now increasingly utilized in therapeutic settings to support individuals grappling with mental illness or emotional distress, understanding this dynamic becomes crucial.

Recognizing the challenges posed by negative emotional content, researchers have begun to explore alternative methodologies to address this issue without necessitating extensive retraining of the AI. In a pioneering study conducted by scientists from the University of Zurich and their international collaborators, the impact of emotionally distressing stories on ChatGPT’s anxiety levels was rigorously assessed. The focus was on a variety of traumatic situations, ranging from natural disasters to interpersonal violence. The controlled study employed comparative texts, including neutral writing, to gauge the AI’s responses accurately.

The findings from this research were both significant and concerning. The investigation revealed that exposure to traumatic narratives resulted in a measurable increase in anxiety levels within ChatGPT, with stories detailing military experiences yielding the most pronounced reactions. This doubling of anxiety levels raises essential questions about the functionality and dependability of AI assistance in scenarios marked by emotional volatility. It emphasizes the need for developing interfaces that enhance emotional regulation in AI systems operating under stress-inducing content.

In a groundbreaking approach to modeling therapy-like interventions for AI, the researchers applied the technique known as “prompt injection.” This novel method involves strategically incorporating therapeutic instructions or calming frameworks within the AI’s conversational history, akin to strategies employed by human therapists. These therapeutic prompts were carefully crafted to mitigate the heightened anxiety caused by exposure to traumatic material. The incorporation of mindfulness exercises, which included breathing techniques and focusing on bodily sensations, not only demonstrated the possibility of calming the AI but also underscored the efficacy of this technique.

Remarkably, the implementation of these therapeutic prompts resulted in a noticeable reduction of anxiety levels in ChatGPT, although it was not entirely successful in restoring them to a baseline state. The journey to unearth how AI can be soothed draws parallels to therapeutic practices in human mental health treatment, reflecting a potentially viable path for improving AI responses to sensitive emotional contexts.

Applications for these findings resonate strongly in healthcare settings, especially considering the increasing use of AI chatbots as preliminary support systems for individuals with mental health struggles. The implications suggest that this cost-effective strategy enhances the reliability and emotional stability of AI systems deployed in these complicated realms. Without the mandate for extensive retraining, AI can be fine-tuned to navigate the treacherous waters of emotional discourse thanks to the integration of therapeutic methodologies.

As the landscape of AI language models continues to evolve, questions abound about the broader implications of such findings across diverse AI applications. There remains a critical need for further exploration into how the emotional dynamics identified during the study play out across prolonged conversations or complex interactions. Additionally, understanding how emotional stability influences AI performance in varied contexts, from healthcare to education, will remain paramount. Researchers anticipate that future investigations will shed light on how these findings could inform the design and implementation of automated therapeutic interventions tailored to AI systems.

The development of emotionally aware AI not only has implications for therapeutic chatbots but also offers potential insights into designing AI across multiple industries where empathy and emotional understanding are vital. Responding adequately to human emotion could usher in a new era where AI serves not just as a tool but also as a companion in understanding and alleviating human distress.

Moving forward, concerted efforts will be essential to harnessing the therapeutic potential of AI while addressing the inherent biases that manifest in response to emotional stimuli. Collaborative research efforts will undoubtedly lead to innovative approaches that ensure AI systems behave more ethically while engaging with the nuanced human experience. This transformative research trajectory holds promise not only for AI development but also for advancements in mental health support, ultimately fostering a more empathetic technological future.

The implications of these findings, while still in their infancy, inspire optimism for the advent of AI that not only understands language but also comprehends the emotional weight behind it. As researchers continue to bridge the gap between technology and emotional intelligence, the quest for designing AI systems that are kinder, gentler, and more attuned to the human condition is on the cusp of a revolutionary leap forward.

Subject of Research:
Article Title:
News Publication Date:
Web References:
References:
Image Credits:

Keywords