The Rise of Clinical Large Language Models
Large language models (LLMs) such as OpenAI’s GPT-4 and Google’s Gemini, built on artificial intelligence (AI), hold immense potential to support, augment, or even eventually automate 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. This article provides a roadmap for the ambitious yet responsible application of clinical LLMs in psychotherapy. We will first offer a technical overview of clinical LLMs. Next, we will discuss the stages of integration of LLMs into psychotherapy, highlighting parallels to the development of autonomous vehicle technology.
We will then explore potential applications of LLMs in clinical care, training, and research, while highlighting areas of risk given the complex nature of psychotherapy. Finally, we will provide recommendations for the responsible development and evaluation of clinical LLMs, which include centering clinical science, involving robust interdisciplinary collaboration, and attending to issues like assessment, risk detection, transparency, and bias.
Understanding Clinical Large Language Models
Language models, or computational models of the probability of sequences of words, have existed for quite some time. The advent of large language models, enabled by a combination of the deep learning technique transformers and increases in computing power, has opened new possibilities. These models are first trained on massive amounts of data using “unsupervised” learning, where the model’s task is to predict a given word in a sequence of words.
The models can then be tailored to a specific task using methods, including prompting with examples or fine-tuning, some of which use no or small amounts of task-specific data. LLMs hold promise for clinical applications because they can parse human language, generate human-like responses, classify/score text, and flexibly adopt conversational styles representative of different theoretical orientations.
For certain use cases, LLMs show a promising ability to conduct tasks or skills needed for psychotherapy, such as conducting assessment, providing psychoeducation, or demonstrating interventions. Yet to date, clinical LLM products and prototypes have not demonstrated anywhere near the level of sophistication required to take the place of psychotherapy. The gap likely exists between simulating therapy skills and implementing them effectively to alleviate patient suffering.
Stages of LLM Integration into Psychotherapy
The integration of LLMs into psychotherapy could be articulated as occurring along a continuum of stages spanning from assistive AI to fully autonomous AI. This continuum can be illustrated by models of AI integration in other fields, such as those used in the autonomous vehicle industry.
Assistive AI (“machine in the loop”) stage: The AI system has no ability to complete the primary tasks on its own, but provides momentary assistance to increase quality or decrease burden on the human. For example, an LLM could assist with collecting information for patient intakes, providing basic psychoeducation, suggesting text edits for providers, or summarizing patient worksheets.
Collaborative AI (“human in the loop”) stage: The AI system aids in the primary tasks, but requires human oversight. For example, an LLM might generate messages containing session content and assignments, which the therapist would review and adapt as needed before sending to the patient.
Fully autonomous AI stage: The AI system performs a full range of clinical skills and interventions in an integrated manner without direct provider oversight. This could involve conducting a comprehensive assessment, selecting an appropriate intervention, and delivering a full course of therapy with no human intervention.
Progression across the stages may not be linear, and human oversight will be required to ensure that applications at greater stages of integration are safe for real-world deployment. Additionally, certain types of interventions, such as those that are more concrete and standardized, may be easier for LLMs to deliver than highly flexible or personalized treatments.
Potential Applications of Clinical LLMs
LLMs have the potential to automate several time-consuming tasks associated with providing psychotherapy, such as using session transcripts to summarize the session for the provider, integrating with electronic health records to aid with clinical documentation and chart reviews, and producing personalized handouts for patients.
LLM applications could also be used to automate the measurement of therapist fidelity to evidence-based practices, which is crucial for the development, testing, dissemination, and implementation of these treatments. LLMs could provide real-time feedback and support on patients’ between-session homework assignments, helping to bridge the gap between sessions and expedite skill development.
Furthermore, LLMs could be used to provide feedback on psychotherapy or peer support sessions, especially for clinicians with less training and experience, and to support supervision for psychotherapists learning new treatments. In the long term, LLMs could be used to conduct comprehensive behavioral healthcare, including assessment, intervention selection, and therapy delivery.
LLMs could also enhance the detection and development of new therapeutic skills and evidence-based practices. By analyzing large datasets of psychotherapy transcripts, LLMs could yield fine-grained insights about what makes existing techniques work best or even identify previously unidentified therapeutic techniques associated with improved clinical outcomes.
Responsible Development and Evaluation of Clinical LLMs
While the potential applications of clinical LLMs are vast, the development and integration of these technologies must be approached with caution and care. Interdisciplinary collaboration between clinical scientists, engineers, and technologists will be crucial to ensure that LLMs are developed and implemented in a responsible and ethical manner.
Key considerations for the responsible development of clinical LLMs include:
Accurate risk detection and mandated reporting: Clinical LLMs must prioritize the identification of suicidal/homicidal ideation, child/elder abuse, and intimate partner violence, and have mechanisms in place to report these risks.
Maintaining “healthy” interactions: LLMs must be trained and monitored to prevent the expression of undesirable behaviors, such as depression or narcissism, which could harm vulnerable patients.
Integrating psychodiagnostic assessment: Clinical LLMs should facilitate appropriate intervention selection and outcome monitoring by incorporating diagnostic capabilities.
Addressing bias and ensuring equity: LLMs must be developed with bias countermeasures to prevent the perpetuation of racism, sexism, and other forms of discrimination.
Promoting transparency and trust: Clinical LLMs should be explicitly labeled as such, and their limitations and capabilities should be communicated clearly to patients and providers.
The evaluation of clinical LLMs should follow a hierarchical framework that prioritizes risk and safety, followed by feasibility, acceptability, and effectiveness. Rigorous head-to-head comparisons with standard treatments will be necessary to provide empirical evidence of the utility of these technologies.
Conclusion: Towards a New Era of Personalized Mental Healthcare
The integration of AI and LLMs into mental healthcare holds immense promise, but also significant challenges. While clinical LLMs could revolutionize the delivery, accessibility, and personalization of psychotherapy, they must be developed and implemented with the utmost care and responsibility.
By centering clinical science, fostering interdisciplinary collaboration, and addressing key ethical and technical considerations, the field can work towards realizing the full potential of these technologies to improve mental health outcomes and access to care. As AI continues to evolve, it is crucial that psychologists, clinicians, and technologists work together to ensure the responsible and effective integration of LLMs into the future of mental healthcare.
