Large language models could change the future of behavioral healthcare: A proposal for responsible development and evaluation

Abstract

Large language models (LLMs) built on artificial intelligence (AI) – such as ChatGPT and GPT-4 – hold immense potential to support, augment, or even replace psychotherapy. Enthusiasm about such applications is mounting in the field as well as industry. These developments promise to address insufficient mental healthcare system capacity and scale individual access to personalized treatments. However, clinical psychology is an uncommonly high stakes application domain for AI systems, as responsible and evidence-based therapy requires nuanced expertise. Here, we provide a roadmap for ambitious yet responsible applications of clinical LLMs. First, we discuss potential applications of clinical LLMs to clinical care, training, and research, emphasizing imminent applications while highlighting areas that present risk given the high-stakes, complex nature of psychotherapy. Second, we outline a continuum of assistive to fully autonomous clinical LLM applications that could be integrated into digital treatment modalities, analogous to the development of autonomous vehicle technology. Third, we outline recommendations for the responsible development of clinical LLMs, which should center clinical science and improvement, involve robust interdisciplinary collaboration, and attend to issues like assessment, risk detection, transparency, and bias. Fourth, we offer recommendations for the critical evaluation of clinical LLMs, arguing that psychologists are uniquely positioned to scope and guide the development and evaluation of clinical LLMs. Last, we outline a vision for how LLMs might allow for a new generation of studies of evidence-based interventions at scale, and how these studies may challenge assumptions about psychotherapy.

Mixed Method Feasibility Study Protocol for Socrates 2.0: A Novel Cognitive Behavioral Therapy-Based Generative AI Tool to Facilitate Socratic Dialogue (Preprint)

“…6 The ability of large language models (LLMs) to mimic human language provides the opportunity to build digital mental health tools that may feel more natural and engaging. 7,8 Building on recent advances relating to LLMs, our goal was to create and examine the feasibility of a generative AI tool that can complement traditional cognitive behavioral therapies (CBTs) by facilitating a core therapeutic intervention: Socratic dialogue. 9,10 Given current limitations of LLMs, 7 our tool "Socrates 2.0" was designed to ultimately be used in conjunction with a licensed clinician to make CBTs' out-of-session practice of evaluating one's thoughts more engaging compared to traditional worksheets.…”

“…We viewed the Socratic process as a good specific use case for an LLM-based tool that guides patients through this process. Given the advanced conversational abilities of LLMs, 7 we hypothesized that a generative AI tool would be able to engage patients in Socratic dialogue at least as well as a newly trained therapist, and that employing LLMs would make out-of-session practice, such as examining one's thoughts, more engaging for patients than worksheets. 21…”

“…Given some of the known issues associated with LLMs, a key goal for our team was to ensure that Socrates 2.0 would not only provide users with a solid Socratic dialogue experience, but also that it would not provide harmful responses. 7 This latter part was particularly important given its intended use in the context of mental health care. Our team's internal testing process…”

Making Moral Decisions With Artificial Agents As Advisors. AnfNIRS Study

“…At home, social chatbots (Henkel et al, 2020;Pentina et al, 2023) are being increasingly used as personal assistants (e.g., Alexa, Siri, Fireflies, Google Assistant), friendship companions (e.g., Replika, Anima, Kajiwoto, Microsoft XiaoIce) or even relational agents for (mental) healthcare (e.g., Woebot; Wysa; Youper; Stade et al, 2023). At work, intelligent decision support systems are now being used as advisors in healthcare, finance, and the military to name a few (Černevičienė & Kabašinskas, 2022;Macey-Dare, 2023;Shortliffe, & Sepúlveda, 2018;Wasilow & Thorpe, 2019).…”

Formulation as Representation - Modelling the Cognitive Space of Mental Health Clinical Reasoning

“…Mental healthcare in particular is an area in which the abilities of LLMs and other forms of AI arguably have the most transformational potential (Hua et al 2024). This is because the degree of cognitive and administrative work required by mental healthcare clinician, particularly when working with the most complex problems, is one of the highest in all areas of healthcare (Stade et al 2023). Previous papers have argued that the use of generative AI in mental health requires particular safety diligence due to the interpersonal nature of mental healthcare (Hider and Wright 2023).…”