AI mental health chatbots feel empathetic, but safety evidence still lags

A new scoping review finds that AI-powered mental health chatbots can feel accessible, personal, and empathetic, but the evidence still leaves key questions about safety, sustained use, and real-world clinical value.

Study: Generative AI mental health chatbots: a scoping review of intervention design and user experience. Image Credit: ZilverlightArt / Shutterstock

In a recent article in press in the journal npj Digital Medicine, researchers reviewed the user experience (UX) and intervention design of generative artificial intelligence (GenAI) mental health chatbots.

Around 25% of people worldwide experience a mental health problem, yet most, approximately 85%, do not receive adequate treatment due to barriers such as stigma, costs, professional shortages, geographic distance, and structural inequities, among others. The growing prevalence of mental disorders, alongside treatment gaps, has led to innovative approaches for mental healthcare. In this context, digital tools have garnered attention for delivering mental health interventions due to their scalability and convenience.

Digital mental health interventions provide treatment or support through chatbots, websites, mobile applications, and wearables, among others. Conversational agents, or chatbots, are applications that simulate human dialog using machine learning and natural language processing algorithms. Traditional mental health chatbots deliver pre-scripted therapeutic content using rules- or retrieval-based systems, but are limited in their ability to personalize support and recognize user needs.

Unlike traditional chatbots, large language model (LLM)-based chatbots can simulate core aspects of therapeutic encounters, including personalized suggestions and empathetic reflections. However, GenAI systems may produce incorrect or inappropriate responses. In addition, the open-ended conversational capacity of GenAI systems makes designing interventions more consequential and complex compared to rules-based systems.

The study

In the present study, researchers reviewed the design and UX outcomes in interventions involving GenAI mental health chatbots. First, a systematic literature search was performed to identify studies on the design and deployment of GenAI mental health chatbots. Reviews, editorials, media articles, and commentaries were excluded. In total, 21 studies conducted in 11 countries from 2023 to 2025 were selected.

The largest numbers of studies came from China and the United Kingdom, followed by the United States. Interventions ranged from early-stage prototype evaluations to clinical trials and one real-world implementation study. Most studies included general or clinical adult populations, including older individuals with dementia, while some involved simulated users and university students. Sample sizes ranged from five to 527 participants. There was substantial heterogeneity in outcome measures across studies, and most interventions remained at an early stage of development.

Intervention design characteristics

Chatbot interventions most often targeted depression and anxiety and adopted shared therapeutic mechanisms, including mindfulness, emotion regulation, and cognitive restructuring. Some systems also focused on mental well-being, stress, and loneliness, and emphasized general support and preventive care rather than specific treatment. Others targeted eating disorders, post-traumatic stress disorder, and dementia.

Dementia-focused interventions aimed to address its related psychological dimensions, for example, caregiver burnout, psychological distress, and loneliness, rather than its central neurological features. Further, most interventions were based on cognitive-behavioral therapy principles, including behavioral activation, psychoeducation, Socratic dialogue, acceptance and commitment therapy, cognitive restructuring, and mindfulness.

Interventions also varied in frequency, delivery modality, and duration, with the majority being short-term, ranging from two to eight weeks. They were mostly deployed via web-based interfaces and mobile applications. Some tools were deployed through messaging or social media platforms. About 67% of interventions were non-embodied text-based chatbots, whereas others used voice, avatar-based, augmented reality, or other multimodal interactions to enhance engagement and realism.

UX of GenAI mental health chatbots

All but two studies evaluated at least one UX domain, with the majority using quantitative measures, for example, Likert scales. Some studies used qualitative feedback, such as open-ended questions and semi-structured interviews. The most common outcomes were user satisfaction and acceptability. Participants described the interventions as convenient and accessible across studies, with generally moderate-to-high acceptability and high reported user satisfaction.

Half of the studies examined usability using qualitative feedback, the System Usability Scale, or Likert scales. Interface design, interaction mode, and deployment platform were observed, along with differences in usability across studies. Users preferred free-flowing chat interfaces and customizable features over predefined options. Further, some interventions had unclear scope or limited functionality, leaving users unsure of the chatbot’s capabilities.

Only some studies reported objective utilization and engagement metrics, such as session frequency, interaction duration, retention over time, and task completion. Attrition patterns often emerged over time in repeated-measures designs. For instance, uptake was initially high in multi-week interventions but declined over time. Further, most chatbots featured personalization, reflecting their ability to adapt conversations and UX interfaces to previous interactions and emotional state.

The most common personalization strategy was emotion detection with adaptive interaction, enabling users to receive tailored interactions and empathetic reflections. Studies did not consistently measure perceived impact as a standalone metric. It was typically incorporated within qualitative feedback or broader UX evaluations. In the intervention with the most granular data, the most common perceived benefit was improved clarity and awareness.

However, the review also found that personalization and empathy did not consistently translate into stronger clinical outcomes or sustained use. Some users reported repetitive, generic, or contextually misaligned responses, while others raised concerns about over-reliance on chatbots, reduced human contact, data privacy, and the ability of systems to respond appropriately during crises. Inaccurate or clinically misaligned outputs were also linked to erosion of trust and disengagement in some studies.

The authors noted that several design features were associated with, rather than proven to cause, better UX outcomes. These included familiar deployment platforms, richer interaction modalities, integration into existing care pathways, personalization, domain knowledge grounding, structured delivery, proactive outreach, and co-design with experts and end users. However, the predominance of early-stage studies and limited direct comparative analyses prevented firm conclusions about which features directly improved UX.

Concluding remarks

Collectively, GenAI chatbots have meaningful potential to deliver tailored, empathetic mental health support and show promising acceptability. Nevertheless, standardizing UX assessment, grounding intervention design in user needs and preferences, and sustaining engagement remain major challenges. Addressing them will require co-design with experts and users, validated UX metrics in long-term studies, transparent reporting standards, independent evaluation, clearer reporting of model design and training data, and stronger attention to safety, equity, and crisis-response limits.