Artificial Intelligence in Psychotherapy: Enhancing Clinical Practice While Preserving the Human Element

Introduction

Generative Artificial Intelligence (GenAI) is transforming psychotherapy and is playing an increasingly significant role in mental health, its primary area of application to date (Cruz González et al., 2025). While AI undoubtedly offers substantial benefits –such as expanding access or supporting personalization of care (Oo et al., 2026)-, it also entails important risks and limitations, as well as ethical dilemmas that must be addressed. According to a 2025 study conducted in the United States, more than 48% of participants (499 individuals aged 18 to 80) reported having used Large Language Models (LLMs)—generative systems that provide responses in text format—for psychological support in the past year (Rousmaniere et al., 2025). Participants reported primarily using LLMs for conditions such as anxiety (73%) and depression (60%). The main reasons users cited for consulting these tools included accessibility (90%) and perceived effectiveness. Sixty three percent reported improvements in their mental health as a result (Rousmaniere et al., 2025).

A systematic review of 85 articles published since February 2024 (Cruz González et al., 2025) indicated that the AI tools studied, primarily machine learning algorithms, showed significantly high levels of effectiveness (greater than 75%) for the detection, classification, and prediction of risk of having a mental health condition. AI tools appeared to be accurate also at predicting treatment response and monitoring progress. Chatbots or conversational agents used for psychological interventions yielded the most inconsistent results among the studies reviewed. While the review explicitly states that further research is needed in the field and raises ethical and methodological concerns, it concludes that the potential of AI will transform the mental health landscape (Cruz González et al., 2025).

The adoption of AI and other digital tools in psychotherapy is shaped by therapists’ and patients’ attitudes toward these technologies, including expectations regarding ease of use, convenience of access, performance, data security and confidentiality, and organizational or institutional acceptance (Bekes et al., 2025; Aafjes-van Doorn, 2025).

Current clinical applications of AI in psychotherapy

Assessment, diagnosis, case formulation and monitoring

The shift toward digital assessment environments reflects the demand for more efficient and precise data collection. Artificial Intelligence directly addresses this need by transforming static digital inputs into dynamic, actionable insights. Through Natural Language Processing (NLP) and machine learning algorithms, AI can analyze behavioral patterns, speech nuances, and semi-structured text in real-time, significantly reducing the administrative burden on clinicians while enhancing diagnostic precision. AI is redefining psychological assessment by offering tools that surpass many limitations of conventional methods. Acting as a type of “clinical assistant,” AI can analyze questionnaires, detect response patterns, and contribute supplemental information to diagnostic interpretation (De la Fuente Tambo & Armayones Ruiz, 2025).

Video and audio recordings analyzed with AI can also allow for real-time assessment of patients without the need of self-reports instruments or questionnaires. AI can be used to develop automatized scales based on LLM ratings instead of human responses, as in the example of DISCOVER program in Germany (Eberhardt et al., 2025). Another example is the COMPASS tool (Computational Mapping of Patient-Therapist Alliance Strategies), which uses LLMs and natural language processing (NLP) to infer the state of the therapeutic alliance directly from natural language sessions transcripts (Lin et al., 2025). COMPASS maps conversational turns and associates them with constructs from the Working Alliance Inventory (WAI), generating scores for tasks, bond, and goals (the three components that make up the therapeutic alliance).

AI technologies can support early diagnosis and improve diagnostic accuracy through the analysis of large and complex data sets. AI algorithms can process diverse data sources—speech patterns, social media behavior, physiological signals, and early indicators of psychosis, depression, or posttraumatic stress disorder (PTSD)—identifying patterns that human clinicians might overlook. For example, in the English health system, the Limbic Access tool has demonstrated effectiveness in streamlining patient intake processes, specifically for the assessment and referral of people with moderate to mild clinical conditions. Consequently, these dynamic yields a favorable cost-benefit equation, ultimately expanding access to mental health services for a larger population in demand (Rollwage et al., 2024).

Natural language processing models can also play a role in assessing issues of multiculturalism in psychotherapy. Kuo and colleagues (2024) developed a program to identify conversations about ethnicity, gender, religion, and other multicultural themes in sessions and to assess therapists’ humility and cultural comfort based on their interventions (Soma et al., 2026). By systematically analyzing therapist–patient interactions, these tools may help identify missed opportunities to address culturally relevant issues and provide structured feedback that supports therapist training, supervision, and the development of multicultural competence.

Personalization of treatments

Personalized treatment approaches have been investigated in psychotherapy for decades (Nye et al., 2023). Research has long sought to determine which treatments are best suited for groups of patients based on their characteristics, preferences, presenting problems, and other factors—such as treatment content or therapist style (Cohen et al., 2021). Personalized interventions have the advantage of allowing the identification of the specific components and mechanisms of change within a given treatment. AI could also play a fundamental role at all stages of treatment. For example, at the beginning of treatment, AI could be used to identify the most effective strategy for a particular patient. AI could also be used throughout treatment by monitoring change processes and guiding adjustments to the intervention (Gómez Penedo et al., 2023; Lutz et al., 2021).

Continuous AI-supported monitoring enables real-time adjustments to treatment plans, facilitating personalized care. AI can detect early signs of deterioration by tracking behavioral and cognitive patterns, predicting mood fluctuations, and assessing relapse risk with increased precision. Furthermore, applications integrating wearable devices can monitor physiological indicators such as heart rate variability to assess changes in overall well-being during treatment (Atzil-Slonim et al., 2024). When paired with AI, these systems leverage machine learning algorithms to analyze these dense physiological data streams, converting raw biometric signals into predictive models that can alert clinicians to early signs of emotional dysregulation or impending relapse. Ultimately, machine learning models can help optimize case formulation by predicting optimal session frequency, therapeutic modalities, guiding pharmacological decisions or even predicting risk of non- response to treatment (Gómez Penedo et al., 2024).

AI-assisted interventions

AI-based conversational agents represent one of the most widely studied and implemented forms of AI intervention in mental health care. Li and colleagues (2023) find in a systematic review that these tools can alleviate psychological distress, especially those incorporating generative AI, multimodal devices, human-voice simulation, or mobile applications, with beneficial effects observed particularly in both clinical and subclinical populations, as well as among older adults.

Several chatbots have shown reductions in symptoms of anxiety and depression, although effects tend to be modest compared with in-person therapy. Drawing on NLP, machine learning, and deep learning techniques, these agents interpret user queries and generate tailored responses. Many integrate principles from cognitive-behavioral therapy (CBT) to deliver interactive, self-guided exercises; or adapt feedback based on users’ emotional states and conversational patterns to simulate empathic engagement. Examples include:

• Psychological guidance, psychoeducation, and symptom management, provided by tools such as Woebot, Wysa, GymBuddy and Elomia (Farzan et al., 2025; Jiang & Yang, 2025; Yeh et al., 2025).
• Support in crisis or hard-to-reach contexts  such as the “Friend” chatbot, a conversational agent tested in Ukraine to assist isolated women in war zones (Spytska, 2025).
• Suicide-risk support, illustrated by Jaspr Health, a tablet-based digital intervention designed for emergency-department settings. Jaspr delivers structured, evidence-informed therapeutic activities to support individuals experiencing acute suicide risk (Dimeff et al., 2024).

Therapist training and supervision

Generative AI may also contribute to therapist training and supervision by fostering deliberate self-reflection, supporting clinical reasoning, and enhancing the quality of psychotherapeutic practice. Qualitative research suggests that AI platforms used for clinical documentation and case management can provide therapists with reassurance, alternative perspectives, and a sense of having a supervisor “in the room,” while users remain aware of their current limitations, including errors and biases (Matthews et al., 2025). Similarly, Sentio Counseling Center has incorporated AI into therapist education and supervision through its Innovation Lab, using it as a complementary resource for clinical reasoning, structured skills practice, feedback, academic support, and deliberate practice (Rousmaniere & Vaz, 2025).

Opportunities for AI use

As mental-health needs continue to grow worldwide, AI-based tools present new avenues for expanding equitable access to care. From scalable digital interventions to enhanced diagnostic support, these technologies may help reduce disparities and support more responsive, flexible treatment delivery. Several opportunities stand out in the current evidence base.

Accessibility and scalability

AI enables 24/7 scalable interventions—particularly valuable in rural or underserved areas where professionals are scarce—and can provide rapid support during crises.

Cost-effectiveness

 AI tools offer affordable or free alternatives to traditional therapy.

Reduction of stigma

 AI-assisted platforms offer anonymity, encouraging individuals to seek help without fear of judgment.

A Latin American example: The Aiglé Foundation

A significant characteristic of most developments mentioned above are taking place in rich countries and for privileged populations. In the disadvantaged socio-economic contexts of the Global South, the relevance of AI-enhanced tools is particularly significant given the region’s structural challenges. These tools can help democratize access to high-quality training, offering scalable, low-cost opportunities for therapists to strengthen core competencies and receive feedback that would otherwise be unavailable in many institutional contexts. By supporting sustained skills development, AI-assisted deliberate practice may contribute to reducing inequities in clinical outcomes and expanding the overall quality of psychotherapy services in the region.

Within this context, the Aiglé Foundation illustrates an institutional approach to the responsible integration of AI in clinical practice and research. Aigle is a non-governmental organization based in Buenos Aires, Argentina, with additional headquarters in Guatemala and Valencia, Spain, dedicated to psychotherapy, professional training, and applied research. Its cognitive-integrative model, developed by Héctor Fernández-Álvarez (1992, 2001), reflects a longstanding commitment to bridging the gap between clinical practice with research (Fernández-Álvarez & Fernández-Álvarez, 2017). Within this framework, routine outcome monitoring (ROM) plays a central role, informing therapeutic dialogue and supporting day-to-day clinical decision-making. Patients and therapists use a digital platform to complete session-by-session assessments and progress questionnaires, creating a continuous feedback loop. By applying machine learning to this platform data, Aigle’s ongoing research projects analyze treatment processes and outcomes to enhance clinical precision. Specifically, these AI tools estimate individualized change curves to identify “off-track” patients at high risk of treatment failure, while institutional projects use descriptive analyses to provide personalized recommendation criteria that optimize treatment tailoring for each patient.

Today, the organization is expanding this tradition by exploring multimodal data sources—such as audio, video, and transcripts—as inputs for emerging AI-based tools, including large language models capable of analyzing session dynamics in real time. Recordings from intake through follow-up serve multiple purposes: they enrich clinical feedback, provide concrete learning material for new therapists, support supervisors through observation-based training and guided self-reflection, and lay the groundwork for the development of AI-assisted training programs in the future. Moreover, the use of virtual patients to train competencies such as clinical exploration or strengthen empathic responding enhances repetitive practice and facilitates training in safe environments.

Barriers, risks and ethical dilemmas

Despite its potential, the effective and responsible integration of AI into mental health requires careful consideration of its limitations and ethical dilemmas. The debate surrounding the use of AI in psychotherapy today centers primarily on the risk of technology replacing the human connection at the heart of clinical practice (Herbener et al., 2025). The scientific community has spoken out on this matter, warning of some evident problems with every day, unsupervised use of freely available AI tools (Frances, 2025). 

Limitations in emotional understanding and relational responsiveness

While AI systems are increasingly capable of generating responses perceived as empathic, it remains uncertain to what extent they can approximate the emotional and interpersonal processes involved in psychotherapy (Rubin et al., 2024). Current models do not possess subjective emotional experience and may be limited in their ability to interpret implicit meanings, contextual subtleties, or cues conveyed through silence, prosody, and nonverbal behavior. These limitations may affect their capacity to engage with some of the more complex relational dimensions of psychotherapy.

Algorithmic biases

The performance of AI systems depends largely on the quality and representativeness of their training data. If these datasets contain gaps or systematic biases, AI-generated recommendations may inadvertently reproduce or amplify them. Although ongoing efforts seek to reduce these risks, algorithmic bias remains an important consideration when using AI to support clinical decision-making (Cross et al. 2024).

Privacy and data security

Mental-health data are highly sensitive, raising concerns about confidentiality, data misuse, and breaches. Rigorous ethical frameworks and informed consent protocols are essential to regulate the proper and protected use of AI tools. Psychological assessment, for example, introduces critical challenges, particularly regarding the use of sensitive patient data (López & Rodríguez, 2024).

Dependence and automation bias

Clinicians may become overly reliant on AI recommendations (automation bias), while some patients may perceive continuous AI-supported monitoring as intrusive or threatening to their privacy, potentially affecting engagement and trust in treatment (Aafjes-van Doorn, 2025; Monteith et al., 2026).

Risk of triggering severe pathology

Interaction with human-like AI systems may trigger or exacerbate conditions such as psychosis, self-harm, or suicidal ideation, particularly in vulnerable populations like children and adolescents. International experts in the field of mental health are even proposing the inclusion of new pathologies associated with the use of AI, such as “AI induced psychosis” in diagnostic classification systems (Frances, 2025). Rousmaniere and colleagues (2025) indicated that 9% of LLM´s users found that their interactions with AI resulted in false, derogatory, offensive, or harmful behavior-enhancing responses.

Limited clinical validity

Many AI models lack validation across diverse populations, and long-term effects remain largely unknown. Much more research and trials are still needed to test the effectiveness of these AI-assisted interventions using rigorous validated methodologies (Cruz González et al., 2025; Soma et al., 2026).

Regulatory challenges

Clear guidelines and legal regulations for responsible AI use in psychotherapy remain underdeveloped, even though several countries and regions are already implementing local regulatory standards (Shumate et al., 2025). It is difficult to know which path to take, with the Hippocratic principle “First do no harm” as the starting point. Psychoeducation should be the shared goal of social actors and institutions -such as schools, public health system, private sector- committed to the emotional, physical and mental health of individuals.

Attitudinal barriers and implementation

AI tools require specific training for therapists and overcoming attitudinal barriers to be implemented, especially in institutional settings (Aafjes-van Doorn, 2025). These barriers often manifest as clinician skepticism regarding an algorithm’s ability to capture the nuance of human empathy, coupled with anxieties about automated performance surveillance and the perceived threat of AI replacing clinical intuition.

Professional recommendations

The World Health Organization (WHO, 2024) as well as the American Psychological Association (APA, 2024, 2025) recommends strict data-protection safeguards, explicit informed consent, and supervised use of AI tools, emphasizing that such tools should support—but never replace—clinical judgment or widen disparities. Its guidelines for the ethical and responsible use of AI in psychotherapy clarify that the responsibility for treatment always rests with the professional and that digital tools should only be used as complement and support, under supervision.

The APA emphasizes that these new technological tools should serve to minimize inequality in access to mental health service, facilitating the inclusion of vulnerable population. It also encourages the development of interdisciplinary research to evaluate the feasibility, efficacy, and safety of AI use in mental health (APA, 2024, 2025).

Future directions: The essential role of the human element

Across the studies reviewed, a consistent conclusion emerges: although AI may enhance clinical practice, it does not replace human relatability. Current evidence suggests that hybrid models integrating AI-based support with human therapeutic care may represent one of the most promising approaches, combining the scalability of digital tools with the clinical judgment and relational capacities of mental health professionals (Chen et al., 2024). In these models, AI can handle supervised, automatable tasks, enabling therapists to focus on the therapeutic bond, clinical reasoning, reflection, and complex cases—competencies that cannot be delegated. AI should thus be understood as a complementary tool that expands professional capacities and improves access and outcomes, rather than a substitute (Babu & Joseph, 2024).

Psychotherapy is a highly complex discipline grounded in the fundamental human encounter with another person. Such encounters involve nonverbal communication, embodiment, and forms of cognition that AI cannot provide—and that many patients need or prefer. The most effective therapists will continue to be those who integrate both science and art: scientific knowledge and the experiential craft that have long defined the essence of psychotherapy.