A complete introduction to evaluate the use of AI in clinical practice

By Michiel van Vliet, MS

When contemplating whether to use Al in practice, it is worthwhile to understand the legal implications of its use, when working with patients. This starts by understanding the different rules that govern various aspects of Al's use, including whether its use needs to be disclosed to the patient, to what extent Al needs to be explained, if informed consent is needed, and whether the professional can be held liable for any mistakes made by Al or following Al's advice. There are already a number of rules in place; however for some parts regulations represent a moving target, as authorities mostly play "catch-up" with the quickly evolving Al-technology. Whether Al-software belongs to the wellness space and or whether it is designated as a medical device is more clear however. The latter is classified according to the level of risk, both in the United States, the United Kingdom as well as Europe, and follows a step-based approach on the level of evidence needed to prove effectiveness and safety of instance.

General-purpose chatbots such as ChatGPT, fall outside medical device regulations, and following a number of accidents with younger people, authorities started to update rules to make these safer to use. Therefore, a number of states in the United States have introduced rules ranging from the mandatory disclosure of talking to a chatbot, age verifications, and protocols for flagging suicidal ideation. For providers who want to use Al in practice, possible liability hinges primarily on whether the use of Al is considered standard of care. There are therefore only a few cases where the provider can be held liable and this furthermore depends on whether a direct causal link between the Al-influenced error and the patient's harm can be established. Whether the use of Al in therapy needs to be disclosed in the United States depends on if it constitutes a material risk, which can be determined both from the physician's and the patient's perspective. In Europe, under General Data Protection Regulation (GDPR) law, processing medical data (e.g., by Al) is prohibited unless the patient gives explicit consent. Depending on the exact situation, it must be disclosed how Al works, how it influences a clinical decision or when it involves deviation from the current standard of care.

By Maia Nahmod, PsyD

Artificial intelligence is no longer a future prospect; it is actively reshaping clinical practice and patients' experiences of care. Patients increasingly engage with chatbots and large language models for emotional support, self-reflection, or companionship—often without their therapists' knowledge and beyond tools explicitly designed for mental health. Simultaneously, clinicians are beginning to use Al to support documentation, reduce administrative burdens, and inform clinical deci Paragraph king. These uses of Al are reshaping how therapeutic relationships are experienced, co-construcieu, and sustained.

This chapter offers clinicians practical, communication-focused guidance for navigating this evolving clinical landscape. It examines how to talk with patients about Al use, establish and revisit boundaries, obtain informed consent, protect data privacy, and maintain trust within the therapeutic alliance.

Drawing on real-world clinical scenarios, the chapter examines common dilemmas and provides applied strategies, structured guides, and decision-making frameworks for integrating Al into clinical practice while upholding ethical principles outlined in previous chapters.

Specific tools introduced include patient-centered communication strategies for discussing Al, approaches to align ethical principles with effective communication, a decision-making model for Al-related clinical cases, and an Al-specific informed consent checklist—each illustrated through clinical vignettes.

Regardless of clinicians' openness to adopting new technologies, this chapter addresses the critical issues of communication, privacy, safety, and consent that Al introduces into contemporary therapeutic practice.

By Laura Garcia, PhD

The growing use of conversational Al in mental health care is transforming access to support, enabling scalable interventions that complement traditional therapy. This chapter examines how Al systems and human psychotherapists can be understood as distinct yet complementary modes of support, each offering unique relational pathways to psychological relief. While human-delivered care remains essential for many forms of healing, Al's ability to simulate trust and emotional responsiveness challenges the notion that human presence is always necessary for therapeutic benefit. Drawing on empirical evidence from clinical trials and meta-analyses, this chapter highlights Al's efficacy in low-intensity settings, where it can foster therapeutic alliances comparable to human care for some users.

At the same time, it highlights Al's limitations in managing complex emotional dynamics or crisis situations.

Rather than framing Al as a substitute for human therapy, this chapter calls for a more nuanced understanding of how the two modalities can coexist. It applies insights from customer experience research to distinguish functional (performance-based) from experiential (relational and affective) attributes of psychotherapy. It also integrates a stable-dynamic framework to examine how these evolve over time: stable elements provide foundations early on, while dynamic elements sustain engagement through ongoing attunement and rupture repair. This distinction allows clinicians to express more clearly the value of human-led care and to support the creation of Al systems that are value-aligned and complementary to therapeutic work.

The chapter is structured around a practical framework with four quadrants, based on functional-experiential and stable-dynamic perspectives: structured foundations (functional, stable), which Al excels at scaling through automated screenings and goal alignment; adaptive tasks (functional, dynamic), where Al supports adaptive techniques like reminders and exercises; relational ground (experiential, stable), emphasizing baseline trust and preferences that humans uniquely tailor; and attunement and synchrony (experiential, dynamic), where human presence is irreplaceable for emotional co-regulation and repair. Each quadrant details human roles, Al implications, evidence of efficacy, and limitations, with callout boxes for practical examples.

Professionals are guided to leverage Al for efficiency in structured, low-risk interventions while preserving human involvement for relational and experiential needs, ensuring ethical integration that expands care without compromising its relational core.

By Michiel van Vliet, MS

Among all the different Al-systems, there are two major classes of apps that most practitioners are familiar with: Al Scribes and (Mental Health) Chatbots. Al Scribes are applications that transform recorded conversations between patient and therapist, into a formatted progress note such as subjective, objective assessment plan (SOAP). While saving time for most professionals, its use also has challenges such as diminished reflective practice while writing a note, making factual errors, and possible use of these notes for training new Al-models. Mental Health chatbots followed considerable evaluation from clunky rule-based bots, to more engaging and interesting generative-Al chatbots. The former allow for much more control of the conversation, however it is somewhat uninteresting to talk to since every utterance needs to be programmed. Generative-Al chatbots however, are much more fascinating to talk to, since their generative nature allows for a variety of words, phrases, and sentences. However, this comes at the cost of a hard to control system, that has difficulty recognizing or responding properly to psychotic or suicidal revelations. Social companion chatbots are popular among teens; however heavy usage may lead to unhealthy attachments to its "user." Especially with the widespread use of ChatGPT for mental health support, practitioners are advised to at least read about (mental health) chatbots, in order to guide patients towards safe and responsible usage.

By Joann Kozyrev, BA

Al technologies are profoundly impacting the way people process information and the way humans learn at nearly all levels of education. This chapter will explore ways that therapists can leverage these new Al tools for education, training, and professional development. Educators recommend thinking about Al-supported education as two broad areas of emphasis: learning about Al and learning with Al.

Much of this book provides the reader with an opportunity to learn about Al, so the chapter instead introduces the essential Al knowledge and competencies a therapist or trainee should learn in order to be sure they are using Al within their boundaries of competence. Because this is a chapter about learning, a short section provides some basic information on how Al tools "learn" through the most common Al training methodologies and some guidance about more or less complex models for a task.

Organized by audience, the chapter first addresses trainee education, followed by professional development and continuing education for the more seasoned professional. Next, the chapter introduces methods for leveraging Al for client education as well as using Al to support research. In each of these four sections, the reader will find practical examples of how Al can help the user to save time, try out new ideas, and stimulate learning, reflection, and growth in one's professional practice.

The emphasis is on equipping the reader with practical ideals and options to choose from.

Most of these examples rely on two of generative Al's key strengths: finding patterns and generating content customized for a particular audience or persona. Each section provides several sample prompts which have been vetted on multiple publicly available LLMs. These prompts model how the reader can interact with LLMs and chatbots for educational purposes and are designed to be customized. Naturally, the reader is frequently reminded to check the output of any generative Al tool for accuracy and lack of bias.

Altogether, the chapter provides the reader with tools and approaches for responsibly integrating Al into mental healthcare in ways that are aligned with the field's commitment to education. By employing these approaches, the therapist can approach the challenge presented by Al with both curiosity and a rigorous commitment to learning as well as to standards of privacy, ethics, and safety.

The goal is to create a dynamic partnership between clinicians and Al tools which will enrich many stages of a therapist's career and practice.

By Rachel Joy Victor, BS

Artificial intelligence is shifting mental health care from bounded clinical interventions toward "always-on" digital interactions. While large language models (LLMs) can convincingly simulate therapeutic conversations, their foundational design prioritizes user engagement over clinical efficacy. This user-centric model often results in Al sycophancy: validating maladaptive behaviors to maintain platform retention. This directly contrasts with the clinical objective of building patient independence and eventually terminating treatment.

Rather than functioning as autonomous therapists, Al is most ethically utilized as a practitioner extension. To address systemic provider shortages across the USA and EU, Al can optimize clinical workflows through passive risk monitoring, intake triage, and analyzing subtle, multimodal affective shifts during patient sessions.

Therapy is expanding beyond text-based chat into multi-sensory, context-aware ecosystems. Future modalities will likely utilize "virtual agents" and "simulations": deploying composite avatars and "digital twins" to facilitate psychologically safe role-play and graded exposure therapy, as well as "ambient ecosystems": cross-correlating patient biometrics (e.g., heart rate) with environmental data (e.g., GPS, social graphs) to trigger real-time interventions via haptic wearables and bioadaptive environments.

Because human emotion, memory, and complex trauma remain fundamentally "non-computable" by sensors alone, the ethical discernment and embodied presence of a human psychologist remain indispensable.

By Michiel van Vliet, MS

Although for many, the word "Al" brings up difficult to understand technology, the basics are quite straightforward to understand. Through this understanding, it is much easier to put both its capabilities and it's limitations into perspective. For instance, generative-Al chatbots write text through predicting the next most likely word in a sentence. This immediately helps explain why these chatbots

"hallucinate" (make up facts), as it is a simple probabilistic mechanism that doesn't understand the difference between fact and fiction. This chapter clarifies how artificial intelligence, or the more technically correct definition machine learning works, without going into too much technical depth.

Although the way that machine learning works can be a bit abstract, the reader is encouraged to at least read the first part: Al's core concepts: data, training, and a model.

The chapter will also cover other key definitions, ranging from large language models, to generative Al, to chatbots. Defining this vocabulary can help clinicians express themselves correctly in discussions about Al. The inner workings of popular software such as ChatGPT will be explained, as well as the reasons for its safety issues and its difficulties with displaying factually truthful information (hallucinations). For those who want to dive deeper, the boxes contain an even more detailed description of these technologies.