From Clinical Judgment to Machine Learning: Rethinking Psychotherapeutic Decision-Making with Artificial Intelligence

According to the World Health Organization (WHO; n.d.) mental health disorders, such as anxiety disorder, bipolar disorder, schizophrenia and post-traumatic stress disorder (PTSD), are some of the most significant public health challenges in the WHO European Region. Within this region (which includes 53 countries across Europe and parts of Central Asia), mental health disorders are the leading cause of disability and the third leading cause of overall disease burden. Among these disorders, depression remains one of the most common mental illnesses globally, yet a staggering 66% of affected individuals continue to live with unmet treatment needs (Eilert et al., 2021; World Health Organization, 2023).

Empirically supported psychotherapeutic treatments have demonstrated strong efficacy, are endorsed by clinical guidelines, and are widely used in mental health care as a preferred first-line treatment option (Lorimer et al., 2021). A substantial body of research and numerous clinical trials have affirmed their effectiveness across a wide spectrum of mental and behavioral health disorders (Eilert et al., 2021), applicable to diverse settings (e.g., primary care medicine, community health, specialty treatment services), and across the lifespan (Nathan & Gorman, 2007). In addition, psychotherapy research has identified both specific and non-specific factors that contribute to treatment outcomes (Norcross & Lambert, 2019). Beyond core techniques and strategies, broader factors, such as the quality of the therapeutic relationship, therapist competence, and adherence to protocols, significantly shape psychotherapy’s clinical effectiveness. As such, psychotherapy is a fundamentally human-centered practice, dependent on the dynamic interplay of targeted interventions delivered within a professional, relational framework to effect meaningful clinical change (Aafjes-van Doorn et al., 2020).

At the same time, artificial intelligence (AI) and machine learning are rapidly advancing and are increasingly applied to the field of mental health, including psychotherapy (Burr & Floridi, 2020; Torous et al., 2020). These technologies aim to assist individuals in learning and applying therapeutic skills, identifying behavioral patterns, and integrating interventions into daily life and by drawing on well-established approaches, such as cognitive behavioral therapy (CBT), positive psychology, and mindfulness (Prescott & Barnes, 2024). Some AI-based conversational agents and chatbots are even designed to simulate emotional intelligence with the goal of forming therapeutic alliances with users, clients, or patients (Darcy et al., 2021; Ghandeharioun et al., 2019).

The most promising applications of computer-assisted interventions (CAI) are seen among vulnerable or underserved populations, such as the elderly, adolescents, and individuals facing stigma, financial limitations, or a preference for non-traditional care (Fiske et al., 2020; Luxton, 2020). Furthermore, AI holds the potential to enhance mental health care through personalized treatment delivery, support for therapeutic technique implementation, empowerment of patients, and early detection of mental health conditions via digital phenotyping (Onah et al., 2024; Tekin, 2020). These potentials are underscored by the growing research interest in CAI and the rising number of developers and service providers in the mental health technology space (Bendig et al., 2022; Fiske et al., 2020).

However, the implementation of CAI also introduces a range of complex challenges and unresolved questions that must be explored to fully understand its long-term societal and individual impacts (Sedlakova & Trachsel, 2023). Even though millions of people now interact with so-called “digital psychotherapists,” the true efficacy and implications of such technologies remain insufficiently evaluated (Shatte et al., 2019). While preliminary studies point to benefits in areas like prevention, treatment, and relapse management, the current evidence base is still emerging and should be interpreted cautiously (Bendig et al., 2022; Thieme et al., 2020; Torous et al., 2020). Compounding this, many widely promoted chatbots and conversational agents lack rigorous empirical validation, creating a wider gap in the literature of this uniquely complex topic (Bendig et al., 2022; Kretzschmar et al., 2019).

Much of the current ethical discourse around AI-driven mental health tools is focused on immediate concerns such as privacy, data security, and the evidentiary foundation of interventions (Bauer et al., 2020; Luxton, 2020; Wang et al., 2020). Although these are important considerations, this narrow lens restricts the capacity to anticipate the broader and more enduring implications of these technologies. AI in mental health intersects with multiple aspects of life and involves diverse stakeholders, thereby requiring a more comprehensive and integrative approach (Rubeis, 2021). Such an approach must consider conceptual, epistemic, normative, and ethical dimensions of human-AI interaction while also grounding analysis in the four core principles of biomedical ethics: Autonomy, Beneficence, Non-maleficence, and Justice.

Nevertheless, even this framework faces limitations in addressing deeper philosophical concerns. As emerging technologies evolve, they may shift the foundational meanings of ethical principles; what constitutes harm, for example, may look different in a digital therapeutic context than in an in-person therapeutic context. Integrity remains a core value in psychotherapy, where deception is only acceptable in rare and justified circumstances (American Psychological Association, 2017). These raises pressing questions in the digital domain, such as whether it is ethically permissible for AI chatbots to simulate empathy, and the literature strongly advocates for more holistic, long-term, and human-centered evaluations that account for the evolving individual and societal impacts of these technologies (Bendig et al., 2022; Burr & Floridi, 2020; Rubeis, 2021). Below is a chart describing differences between traditional psychotherapy and artificial intelligence psychotherapy (Spytska, 2025).

---

Aspect	Traditional Therapy	AI Therapy (e.g., Chatbots)
Personal Interaction	Direct, face-to-face with a therapist	No direct human interaction, virtual conversation
Emotional support	High level of personalized, empathetic care	Limited emotional support; relies on programmed responses
Accessibility	Limited by therapist availability and geographic location	Highly accessible, available 24/7 regardless of location
Scalability	Limited by number of therapists and session time	Highly scalable, can serve many users simultaneously
Adaptability	Therapists can adapt care based on real-time feedback	Adapts responses based on user input but lacks human nuance
Crisis Situations	Highly effective, but may be inaccessible in crisis zones	Easily accessible in crisis zones with internet access
Effectiveness in reducing anxiety	Higher effectiveness due to personalized care	Effective but less than traditional therapy
Cost	Typically, more expensive due to human resources	More cost-effective due to automation

---

Source: Spytska (2025).

Thus, from the foregoing, researchers have also emphasized the need for ethical guidelines and clearly defined criteria for the development and application of AI technologies, as well as the importance of equipping health care professionals with appropriate training. Additionally, experts in the ethics of AI in mental health are recognizing the need to explore how interactions with AI may influence core aspects of psychotherapy, such as the therapeutic relationship, individual self-understanding, and personal identity (Burr & Floridi, 2020; Fiske et al., 2020).

Promise, Risks, and the Ethics of Data-Driven Treatment

Digital treatments have introduced a transformative development in psychotherapy: access to multi-dimensional data sets. These data sets containing biological, demographic, clinical, and behavioral variables created a shift beyond traditional inferential paradigms, such as reliance on p values and effect sizes, which have been criticized for limited replicability and clinical utility. When integrated with advances in artificial intelligence and machine learning, such data hold considerable promise for enhancing diagnostic precision, predicting prognosis, and optimizing treatment planning for individuals with mental health conditions, thereby improving overall treatment outcomes (Dwyer et al., 2018).

Clinically, these innovations could lead to more accurate assessments, better patient stratification, and the delivery of tailored or combined interventions (Dwyer et al., 2018). In systems strained by long waitlists and workforce shortages, even modest gains in treatment responsiveness could yield significant benefits at individual, economic, and societal levels (Prasad et al., 2023). Machine learning has been applied to diverse data types, such as session transcripts, patient-reported outcomes, neuroimaging, biomarkers, and sensor-based metrics, and utilized to predict symptom remission, treatment response, relapse risk, and the long-term course of mental disorders (Aafjes-van Doorn et al., 2020; Angstman et al., 2017; Hilbert et al., 2020; Kautzky et al., 2018; Lorimer et al., 2021; Maarsingh et al., 2018; Prasad et al., 2023; Thieme et al., 2020). While the field remains nascent, its convergence with AI provides a strong proof of concept, or evidence of its value and potential, to warrant continued research and innovation (Aafjes-van Doorn et al., 2020; Chekroud et al., 2021; Thieme et al., 2020).

However, the growing enthusiasm for AI in psychotherapy has also led to heightened awareness of its limitations and associated misconceptions (Chekroud et al., 2024). As the volume and complexity of collected data expand, researchers are increasingly challenging the presumed objectivity and reliability of algorithmic inferences. A key concern is the tendency to follow algorithmic recommendations uncritically, without understanding their underlying assumptions and constraints.

Chekroud et al. (2024) highlighted how a machine learning model failed to generalize treatment outcomes for schizophrenia, thereby exposing the limits of algorithmic rationality. Similarly, other studies document instances where technically functional algorithms have caused harm to vulnerable and marginalized groups (Broussard, 2018; Eubanks, 2019). A striking example is the legal sentencing algorithm that disproportionately labelled Black individuals as high-risk for recidivism, while incorrectly assessing White individuals as low-risk, resulting in racially-biased sentencing (Eubanks, 2019).

Biases, heuristics, arbitrary classifications, and noisy data embedded in today’s multi-dimensional datasets introduce additional uncertainty and error (Hong, 2021). These systemic issues underscore the dual reality of AI in psychotherapy. While its potential is compelling, its ethical, social, and epistemological counterpoints warrant critical scrutiny. Psychotherapy data, like sentencing data, often reflect historically biased systems. For instance, cognitive behavioral therapy (CBT), a dominant modality in the field of psychotherapy, has been developed and validated primarily with White, well-educated, heterosexual populations. As a result, algorithms trained on such data may fail to account for or appropriately represent neurodiverse individuals, racial and ethnic minorities, LGBTQ+ groups, culturally diverse populations, and people from various socioeconomic backgrounds (Hilbert et al., 2020).

The contextual and lived experiences that are central to effective psychotherapy are frequently underrepresented in existing datasets. In the case of the schizophrenia study mentioned above, datasets were chosen based on their comparability and consistency; traits suitable for algorithmic processing, but not necessarily for reflecting complex, real-world mental health experiences (Wong, 2023). This methodological trade-off often results in the exclusion of environmental and cultural factors critical to treatment outcomes. Consequently, AI applications in psychotherapy risk perpetuating homogeneity, standardizing care, and rendering the nuanced realities of diverse clients invisible (Crawford, 2021; Wong, 2023).

Despite these limitations, some recent studies highlight the perceived benefits of AI-based mental health tools. For example, Sweeney et al. (2021) surveyed 100 mental health professionals across five countries and found that more than half viewed chatbots as helpful in supporting clients’ health management. Similarly, the Wysa app study demonstrated that participants who engaged with the app showed improvement in depression symptoms, especially high-engagement users, and over half of the users described the app as helpful and encouraging (Inkster et al., 2018; Palacios et al., 2022; Santomauro et al., 2021). Building on this, researchers are increasingly exploring ways to aggregate diverse psychotherapy databases to enhance model accuracy (Chekroud et al., 2024; Ulberg et al., 2023). However, data completeness remains a challenge, spurring debates around datafication, the process of transforming various aspects of human life, behavior, and experiences into quantifiable digital data. Some scholars argue for strict limits due to privacy risks, while others advocate for innovation through privacy-enhancing technologies that enable the secure integration of varied data sources (The Royal Society, 2023). Additionally, synthetic data (artificially-generated datasets) offer a promising alternative. These datasets are customizable, cost-effective, easy to produce, and inherently protect privacy. Most importantly, they can be designed to include underrepresented groups, thereby addressing inclusivity gaps that exist in many real-world datasets.

Recommendations and Future Perspectives

AI systems used in psychotherapy are only as good as the data and assumptions behind them. Many machine learning models are trained on datasets that lack cultural diversity or overrepresent certain groups—often White, Western, and socioeconomically privileged populations (Kautzky et al., 2018; Swaminathan et al., 2023). This creates a risk that the algorithms will reflect and perpetuate existing cultural biases and stereotypes. For clinicians, it is crucial to approach AI tools with a critical mindset: understand where the data comes from, which populations are included or excluded, and how the algorithm’s recommendations are generated. Blind reliance on these tools can lead to misdiagnoses (Wang et al., 2020), inappropriate treatment plans, and overlooking culturally specific symptoms and experiences (Onah, 2024b). By questioning and understanding the limitations of AI, psychotherapists can better safeguard their clients from harm and ensure that the care they provide remains equitable and respectful.

AI can offer valuable insights, but it cannot replace the nuanced, holistic understanding that a skilled therapist and clinician brings to each session. Clients’ cultural backgrounds, life histories, social identities, and unique contexts play a pivotal role in their mental health and how they experience symptoms and respond to treatment (Biswas & Talukdar, 2024). Therefore, therapists must interpret AI-generated suggestions in light of these factors rather than following them unquestioningly. This means actively integrating the client’s voice and narrative, assessing how cultural norms, stigma, or socio-political realities shape their mental health journey, and adapting interventions accordingly. AI tools should be seen as adjuncts that support, rather than substitute, culturally responsive clinical judgment (Shatte et al., 2019).

Also, clinicians and psychotherapists have a role to play beyond the therapy room. They can advocate for transparency and inclusivity in the development of AI tools by engaging with researchers, developers, and policymakers. This involves demanding that datasets used to train AI models incorporate a broad spectrum of populations, including racial and ethnic minorities, LGBTQ+ communities, neurodiverse individuals (Mosquera et al., 2024) and people from various socioeconomic backgrounds, to ensure that AI recommendations are relevant and fair across diverse groups (Roa et al., 2021). Clinicians can contribute to or support efforts that audit and evaluate AI for bias and advocate for ethical standards that require accountability from AI developers. Such advocacy helps to move the field toward technologies that truly support all clients and reduce health disparities.

As AI becomes more prevalent in psychotherapy, professional education and training must evolve. It is not enough for clinicians to be digitally competent; they must also develop cultural humility and sensitivity specific to AI applications (Thieme et al., 2020). This includes understanding how systemic inequalities and historical injustices might be encoded in algorithms, how bias manifests in AI outputs, and how these biases can affect clinical outcomes (Onah, 2024a). Training programs should provide frameworks for critically evaluating AI tools, recognizing potential harms, and implementing strategies to mitigate bias. By integrating cultural awareness with digital literacy, clinicians can better protect vulnerable populations and provide care that respects diverse experiences (Sedlakova & Trachsel, 2023).

At its core, psychotherapy is a deeply human endeavor grounded in empathy, trust, and relational understanding, all qualities AI cannot authentically replicate. While AI can help identify patterns, track progress, or suggest interventions, it cannot grasp the complexity of human emotions or the lived reality of clients (Eilert et al., 2022). Psychotherapists should use AI tools as supportive instruments rather than definitive authorities. Maintaining a reflective, client-centered stance means continually questioning whether AI-driven recommendations align with the client’s values, experiences, and goals. This approach ensures that technology enhances rather than undermines the therapeutic alliance and the ethical foundation of psychotherapy (Eilert et al., 2022).

In conclusion, AI’s capacity to analyze multi-dimensional data offers significant potential to enhance diagnosis, personalize treatment, and improve outcomes in mental health care. However, concerns about bias, ethical issues, and algorithmic limitations persist, especially regarding the exclusion of diverse populations and loss of contextual understanding. To responsibly integrate AI, clinicians must prioritize cultural sensitivity, uphold empathy, and critically assess AI tools. Ethical principles like autonomy, beneficence, and justice must guide this integration. Advocacy for inclusive data and interdisciplinary collaboration is vital. Ultimately, AI should support (not replace) human clinical judgment, requiring ongoing research and reflection to ensure equitable, effective psychotherapy.