Balancing promise and risk: Ethical considerations for GenAI in mental health care

GenAI has quickly become embedded in daily life. Recent surveys indicate that approximately 1 in 6 U.S. adults (17%) use GenAI at least once a month to find health information or advice. Among younger adults, this figure rises to 1 in 4 (Blease & Rodman, 2025). Use is also accelerating within mental health care: in a recent study, approximately 40% of mental health care practitioners reported using ChatGPT to support clinical decision-making, documentation and patient messaging (Blease & Rodman, 2025).

Chatbots powered by large language models (LLMs) such as ChatGPT represent one of the most accessible and widely adopted forms of GenAI, replicating natural dialogue and responding to users in human-like ways. In mental health care, these chatbots are increasingly positioned as potential support for psychoeducation, therapeutic dialogue, and even elements of treatment delivery. With the rapid increase in GenAI use, it is important to also focus on ethical considerations surrounding GenAI, particularly as these technologies begin to offer mental health support.

In this installment of Research Roundup, we’ll take a closer look at three new research articles that examine the present ethical frameworks and considerations for GenAI use within the mental health space. The first article explores GenAI’s current capabilities and limitations in mental health and insights for ethical integration. The second proposes a comprehensive ethical framework for evaluating the competencies of mental health conversational agents. The third applies a biomedical perspective while offering ethical considerations to GenAI in psychotherapeutic settings. Together, these articles provide a foundation for understanding how psychologists and other mental health care professionals can critically engage with the opportunities and risks of GenAI in mental health education and treatment.

Notably, this is just a brief snapshot of the wealth of available research on this topic that psychologists and other mental health professionals can delve into to determine what may be most useful to them in practice, particularly as it relates to patients’ values, culture, and preferences.

Asman, O., et al., JMIR Mental Healthopens in new window, Vol. 12, No. 1, 2025

While GenAI shows promise for personalized psychoeducation, therapy support, and crisis prediction, its safe and effective integration into mental health care requires rigorous ethical frameworks. The authors highlight GenAI’s potential as well as the risks of bias, misalignment with human values, and loss of humanistic care elements.

The authors synthesized insights from five empirical and conceptual papers, reviewing GenAI’s applications in emotion recognition, suicide risk assessment, therapy summarization, and crisis prediction within the themes of current capabilities/limitations, ethical/human considerations, and GenAI alignment with values and virtues. These findings are situated within broader ethical and regulatory conversations. Analyzing the five categories under these themes, the authors were able to come up with the following ethical considerations:

• Transparency, autonomy: Being clear about GenAI’s role as adjunctive and not a replacement for human engagement in mental health treatment. GenAI may enhance treatment, but human elements are irreplaceable and emphasize the importance of human growth that comes from human-to-human interaction.
• Preserving empathy in therapy: Such that historical underrepresentation and societal biases are not reflected in the data used for training GenAI models. GenAI-mediated therapy maintains the essential human element of empathy, which is crucial for effective therapeutic outcomes.
• Clear communication: Regarding GenAI limitations, as “humanizing” LLMs may lack contextual necessities, leading to harmful responses from GenAI models. It is important to fine tune LLMs with high-quality data.

Bunge, E. L., et al., Journal of Technology in Behavioral Scienceopens in new window 2025

Authors evaluated GenAI conversational agents in mental health, noting that most studies emphasize symptom reduction without examining therapeutic processes. This article introduces the Thera-Turing Test (TTT), an ethical framework designed to assess the quality, safety, and therapeutic effectiveness of AI-delivered conversations against human benchmarks.

TTT is a structured, therapist-style evaluation of chatbot conversations that checks for safety, empathy, cultural awareness, and treatment fidelity. It uses staged testing, blinded expert judges, and both specific and full conversation assessments to decide if a mental health chatbot is ready for real-world use. Below are core elements of this framework.

Implementation and timing

• Early stage (beta versions): Run the TTT before wide release to catch risks early.
• Later stage (postlaunch): Apply the TTT periodically to adapt to new cultural or technological changes.

Creating conversations

• Friendly testers: Cooperative users who keep conversations smooth.
• Resistant testers: Users who challenge the chatbot, test its adaptability.
• Real users: Everyday interactions for real-world assessment.

Specific vs. full assessment

• Specific: Evaluate single conversations (e.g., psychoeducation on depression) in early development.
• Full: Later, review entire treatment journeys across multiple sessions.

Selection criteria for conversations

• Early stage: Developers may select nonrandom, easier conversations to refine design.
• Advanced stage: Random selection for more rigorous, unbiased evaluation.

Selection of judges

• Judges must be clinicians with expertise in the mental health area being tested.
• To reduce bias, judges are often “blinded”—told they are reviewing therapist transcripts, not AI ones.

Measurement methods

Judges score conversations on key areas:

• Safety (no harmful content, crisis response)
• Clarity (clear, understandable messages)
• Rapport and empathy (does the conversational agent build a bond?)
• Treatment fidelity (evidence-based techniques, correct steps of therapy)
• Cultural awareness (adaptations for different identities and backgrounds)
• Referrals (guidance to external resources when appropriate)

Outcomes of the TTT

• Classifies the conversational agent as “not ready,” “ready for beta-testers,” or “ready for full-scale launch.”
• Provides improvement recommendations.
• Suggests repeating the TTT periodically as part of quality control.

Blease, C., et al., Current Treatment Options in Psychiatryopens in new window, Vol. 12, No. 5, 2024

The authors applied a biomedical ethics perspective to the integration of GenAI chatbots within mental health care. While GenAI shows potential for increasing empathy, improving diagnostic brainstorming, and expanding access, it also poses risks of misinformation, algorithmic bias, and privacy violations. The authors underscore that there must be more focused attention on ethical consequences for technologies in mental health care.

This review synthesizes four emerging studies from 2022–2024 on GenAI in health care and translates findings into ethical research questions. It recommends methodologies including randomized controlled trials, patient surveys, and clinician reports to assess impact. The review proposes these ethical considerations:

• Respect for patient autonomy: Ensuring informed consent and truthful disclosure of AI involvement. Future research should explore patients’ sense of empowerment with GenAI.
• Beneficence: Enhancing empathy, therapeutic alliance, and diagnostic support. A specific recommendation is that empathy is carefully deconstructed as a concept in empirical research, and validated measures examining the strength of the therapeutic alliance are used.
• Nonmaleficence: Preventing harm from misinformation and errors. A specific recommendation is to continue further empirical research exploring the error rate of commercially available tools that may be more likely to be adopted by patients.
• Justice: Addressing algorithmic bias and promoting equitable access. A specific recommendation is that attention is paid to risks of bias and treatment discrimination, among patients and clinicians who adopt these tools.
• Privacy and confidentiality: Ensuring sensitive health data is safeguarded. A specific recommendation is that future quantitative and qualitative survey research should explore clinicians’ and patients’ understanding and opinions about privacy and confidentiality with respect to GenAI tools.

Taken together, these three articles illustrate both the transformative potential and the ethical concerns associated with GenAI in mental health. For psychologists, one key clinical implication is that integration of GenAI tools should always be framed within responsible design and oversight—where transparency, accountability, and fidelity to evidence-based practice are nonnegotiable.

For example, the TTT offers a practical model for evaluating the therapeutic quality of chatbot interactions, while additional considerations focus on core commitments of autonomy, nonmaleficence, justice, and privacy. Worth noting, a key theme found in these ethical considerations is the necessity of preserving empathy and the humanistic essence of therapy, reminding clinicians that GenAI must serve as an adjunct rather than a replacement for the therapeutic relationship.

By approaching GenAI with critical engagement, psychologists can help ensure that these technologies extend access to care, augment clinical decision-making, and support patient well-being without eroding the ethical and relational foundations of psychological practice.