The Therapeutic Frontier: Artificial Intelligence in Mental Health Care

Written by Manas K. Deb, PhD, MBA – Co-Founder, MI4People gGmbH 

The problem of mental disorders is quite severe and widespread, with profound impacts on individuals, families, communities, and global economies. Per recent statistics from WHO, globally, about a billion people, or about 1 in 8 individuals, are living with a mental disorder, with anxiety and depression being the most prevalent, closely followed by eating disorders. Mental disorders significantly impair an individual's daily functioning, physical health, and lower life expectancy by as much as 10-20 years. Substance abuse, homelessness, suicide, and even certain chronic diseases like heart disease and diabetes are also strongly correlated to mental disorders. Loss of productivity just from anxiety and depression has been estimated to be over a trillion dollars annually.

Despite the dire impact of mental disorders on human life and the economy, mental health services are underfunded, and access to care is quite inadequate – more than 70% of mental disorders, many of which start in early teens, go undetected or untreated. Even in rich countries like the US, there is only one caregiver for several hundred affected individuals; in poor countries, the situation is dreadful. To make matters worse, stigma and discrimination still discourage many from seeking help. Turning to advanced technology for help in coping with mental disorders is, therefore, only natural. An April 2025 Harvard Business Review article suggests therapy as a major Gen AI use case in 2025.

Integrating artificial intelligence into mental health services represents one of the most promising technological developments in psychological care. As demand for mental health services continues to outpace provider availability, AI technologies are emerging as potential solutions to bridge critical gaps in accessibility and treatment efficacy. Current research indicates that thoughtfully designed AI systems can effectively complement traditional therapeutic approaches while addressing some of healthcare's most pressing challenges.  

Evolution of AI and Machine Intelligence in Mental Disorder Therapy  

The exploration of the use of machine intelligence in medical care for physical ailments started over half a century ago with rule-based expert systems, such as MYCIN and INTERNIST. The emergence of systems tailored to mental health, such as ELIZA in the 1960s, demonstrated the potential of conversational agents in therapeutic contexts, albeit in a rudimentary form. Modern AI has since advanced substantially, integrating machine learning and natural language processing to analyze vast datasets, detect patterns, and support clinical decision-making. Such systems can now assist clinicians by identifying early signs of depression, anxiety, or PTSD through speech and text analysis. For example, machine learning algorithms can detect depressive symptoms by analyzing voice tone, speech patterns, and facial expressions, enabling earlier and more accurate interventions. Furthermore, mobile health applications leverage AI to offer personalized cognitive behavioral therapy (CBT), increasing accessibility to mental health support.  

The Debut of “Therabot” (And, The Behind-The-Scenes Story)   

A publication in the “New England Journal of Medicine AI” on the 27th of March of this year reported significantly positive outcomes from the first ever clinical trial of a generative AI powered tool, named “Therabot”, created by a team led by psychiatric researchers and psychologists at the Geisel School of Medicine at Dartmouth College. The trial involved individuals diagnosed with major depression, anxiety, or some eating disorders. Study participants interacted with Therabot via a conversational smartphone app – talking to it when they felt the need and responding to prompts from the app. The study result showed strong benefits in reduction in symptoms and improvement in mood and well-being: over 50% in the case of depression, over 30% for anxiety, and close to 20% for more difficult-to-treat eating disorders. The researchers behind Therabot concluded that the performance of Therabot was comparable to human therapists. Additionally, they noted that the participants found the app ‘friendly and trustworthy’ like good human therapists.

An account of the training strategy and the large amount of effort that went into the development of Therabot is as interesting as its reported success in therapeutic intervention. Starting way back in 2019, and with the first generation of generative AI technology, the underlying AI model of the app was trained using content available on the Internet, as is commonly done for the foundational generative AI model. The resulting performance was quite unreliable and unsatisfactory, with often unhelpful and ambiguous responses to patient interactions. In the later years, the researchers trained the AI model with information derived from Cognitive Behavioral Therapy (CBT), which resulted in the reported stellar performance. This strategy parallels the training of domain-specific AI models, for example, the ‘physics-aware’ AI systems for solving engineering problems.

High Potential of Broad Help from AI In Mental Disorder Treatment

Beyond direct therapeutic interaction, AI is being used to assist in diagnosing mental health disorders. Machine learning algorithms can analyze data from voice, text, facial expressions, and physiological signals to detect markers of psychological distress. For instance, a 2018 study in Nature Partner Journals: Digital Medicine demonstrated that an AI model could predict the onset of psychosis with over 90% accuracy by analyzing speech patterns. Furthermore, AI is being applied to predict suicide risk. Researchers at Vanderbilt University developed a machine learning algorithm capable of identifying individuals at high risk of suicide up to two years in advance, using electronic health record data. This predictive capability can enable early intervention strategies that are otherwise difficult to implement.

In clinical settings, AI tools have begun to assist therapists with administrative tasks and clinical decision support. For instance, computer vision systems can now analyze facial expressions and vocal patterns during therapy sessions to provide therapists with additional data points regarding client emotional states. For example, the AffectAura system, developed by Microsoft Research, exemplifies this approach by tracking emotional signals to create visualizations that both clients and therapists can use to identify patterns.

We, at MI4People gGmbH, in partnership with Therapieverbund Ludwigsmühle, a German non-profit organization providing help for the addicted, have embarked on a data-driven and AI-powered project called HOPE (see: https://www.mi4people.org/hope) whose goals include analysis of data related to addiction patients and support therapists in creation of treatment plans with the highest chance of success as well as to support help seekers in finding therapy facilities best fitting to their needs. The expected impact of HOPE includes the delivery of effective care with higher efficiency, reducing dropouts, and reducing the burden on therapy counsellors and practitioners.

A Word of Caution 

AI/GenAI is full of capabilities and potential to make a major impact in dealing with mental disorders. Given an estimated half-a-trillion-dollar global market size of related therapies and the present-day ease of creating AI/GenAI apps, many tools are appearing in the market, both from established healthcare providers and health startups. However, this ‘gold-rush’ presents a wide variety of serious concerns, especially since the target user groups are made up of vulnerable humans:

Quality – Many apps appear to be superficial – simple wrappers on top of basic foundational AI models trained on content found on the Internet and are not validated using rigorous scientific methods (see related comments above on the training of Therabot); without proper guardrails, due to ‘hallucinations’ and misinterpretations, AI/GenAI apps can produce erroneous responses that could severely harm patients

Data Privacy and Security – Mental health data is among the most sensitive types of personal information. The collection, storage, and use of such data by AI systems raises serious ethical concerns regarding privacy, informed consent, and data security. There is a risk that data breaches or misuse could lead to harm or discrimination – this necessitates a robust data protection and governance framework. While certain regulations exist, e.g., HIPAA in the US and GDPR in the EU, as well as guidelines from the American Psychological Association (APA), their proper implementation is critical for the safe use of AI/GenAI in the mental healthcare context.

Bias and Ethical Considerations – Lack of diversity in training data leads to biased behavior by AI models, potentially misdiagnosing, discriminating, or neglecting underrepresented populations, leading to ethical concerns. Alongside improving the training data quality, data provenance transparency and bias auditing can help increase fairness.

Risk of ‘Dehumanization’ Of Treatment – The ‘human element’ has been critical for successful patient-therapist interactions. While AI can process data and provide insights, at its current maturity, it lacks the empathy, intuition, and nuanced understanding that human therapists bring to sessions. Thus, over-reliance on AI tools can lead to ‘dehumanization’ of care with its obvious negative impacts.

While the challenges as cited above are undeniably obstacles in the use of AI/GenAI in the mental healthcare therapy context, many of the remedies are under human control. Proper design and usage of AI/GenAI and a ‘human-in-the-loop’ approach are key to overcoming these obstacles. A 2023 article from the Human-Oriented Artificial Intelligence center at Stanford University, titled “A Blueprint for Using AI in Psychotherapy” advocated stage-wise integration of AI into psychotherapy responsibly – this approach is like autonomous vehicle development, staring with AI assisting the therapist with concrete well-defined tasks, then collaborating with the therapist in suggesting therapy actions, and finally graduating to full independence handling the complete lifecycle of clinical and administrative activities.

In Closing

With AI/GenAI capabilities in human-like conversations, handling ambiguity, rationalizing responses, possible incorporation of a wide variety of AR/VR features, and utilization of techniques like differential privacy and federated learning can also help protect user data, the potential of these technologies in helping with mental disorder therapies is enormous. These capabilities are further enhanced by derivative technologies like ‘agentic AI’ capable of completing complex tasks ‘intelligently’. While full-proof autonomous AI/GenAI tools may take some time to be developed and adopted, with the current technology maturity coupled with appropriate human guidance and oversight, AI/GenAI can already be extremely effective in improving the current challenges of providing mental disorder therapy.

With heartfelt gratitude, 

MI4People Team