There are a variety of task-sharing approaches utilized in different contexts. We include general background information in this chapter to capture some of the ways task sharing is implemented; however, we recognize that the use cases outlined do not reflect the reality of all countries or contexts. We hope that these materials can help identify opportunities for additional testing and validation of AI in other types of task-sharing programs across diverse contexts.
An ever-growing shortage of mental health providers exists globally, despite their services being more necessary than ever. More than half of all people will experience a mental health condition in their lifetimes, but there are only 13 mental health providers for every 100,000 people worldwide. Access is even more limited for people in low- and middle-income countries (LMICs), for people in conflict settings, for people with marginalized identities, and for or those in rural and low-income communities.
To address this gap, a spectrum of mental health skill-building programs has been developed to train individuals to provide support or care to those in need. These efforts include training volunteers and laypeople to provide care and support and upskilling licensed professionals. For example, there are programs in which individuals with lived experience are trained to provide peer support (that is, peer support programs), programs in which licensed counselors provide support to individuals in crisis (that is, crisis counseling), and programs in which laypeople, nonspecialist health professionals, or community workers are trained to provide evidence-based care (that is, task-sharing programs).
Since task-sharing programs focus on addressing workforce shortages by expanding access to evidence-based care, the following chapters and sections in this field guide focus on these programs and how AI could scale their efforts. At the same time, we acknowledge that many of the AI use cases and topics explored in this field guide may also apply across the broader spectrum of mental health skill-building programs (such as programs that upskill mental health professionals).
Task sharing is an evidence-based solution that increases access to care. In this approach, specialist healthcare professionals delegate specific tasks to trained nonspecialist providers (such as community health workers, auxiliary health staff, and community members) to deliver direct mental health services to the public. This strategy has also been successfully used to deliver care related to maternal health, HIV, and non-communicable diseases.
There is also strong evidence that task sharing is effective in the mental health domain. Lay providers have been trained in multiple aspects of mental healthcare—from assessment, triage, and engagement to treatment. Empirically supported treatments such as cognitive behavioral therapy (CBT) and interpersonal psychotherapy, for example, have been successfully adapted for delivery by non-specialist providers by nonprofit organizations that operate across different contexts. Task sharing in the mental health space has been shown to improve clinical outcomes, reduce costs, and extend the reach of a limited mental health clinical workforce. Research finds that mental health task sharing is cost-effective and can both increase the number of people treated for mental health concerns and reduce their mental health symptoms. There also is evidence that task sharing for mental health service delivery is effective in low-resource settings such as LMICs.
Several mental health task sharing programs have implemented evidence-based methods across different contexts. However, getting lay providers fully trained and confident, maintaining quality (that is, ensuring providers adhere to the program’s protocols and evidence-based practices when delivering services), and ensuring sustainability (that is, achieving meaningful scale through long-term financial support) requires structured processes and support of many different kinds.
AI offers novel opportunities that can address implementers’ challenges while maintaining quality and sustainability and enhancing efficiencies across the implementation cycle of task-sharing programs. By streamlining training, supervision, and decision-making, AI could help task-sharing programs scale up evidence-based interventions while maintaining high standards of model fidelity. Other mental health skill-building programs not focused on task-sharing might also find these AI solutions applicable to their contexts, enabling them to expand their reach and impact.
What is AI?
AI refers to the capability of computer systems to perform complex tasks typically associated with human intelligence, such as learning, reasoning, problem-solving, decision-making, and creating. We recognize that this is a field that is seeing multiple innovations at a fast pace, and that definitions and guidance related to it may rapidly evolve and change.
Gen AI is a subset of AI that focuses on generating new content, including text, images, music, and other media. It uses machine learning, deep learning, and natural language processing to generate outputs that resemble human-created content. Large language models (LLMs) exemplify how gen AI can produce humanlike content by predicting and assembling words in context. For simplicity, the term “AI” is used throughout this field guide and refers to these tools broadly (Exhibit 1).
Exhibit 1: Subsets of Artificial Intelligence
- Machine learning: A subset of artificial intelligence that focuses on computer algorithms that automatically learn and improve through experience without being explicitly programmed.
- Deep learning: A type of machine learning method that uses neural networks (that is, computation models) with multiple processing layers to learn complex patterns in data.
- Natural language processing: A field of AI that focuses on providing machines with the ability to understand and communicate using written and spoken human languages.
- Large language models: A type of deep learning model trained on massive text datasets to understand and generate natural language—or human-like text—often used in applications such as chatbots and summarization.
How does an LLM work?
At their core, LLMs are models trained to predict the next word in language sequences or “fill in the blanks.”
Transformers allow the model LLMs to weigh the importance of different parts of a text and capture context and meaning. For example, as shown in the exhibit below, the transformer model analyzes the sentence “The red fire engine drove quickly down the” (Exhibit 2). It uses attention mechanisms to connect words that are relevant to each other. This process helps the model understand the relationships between words and generate a more coherent and meaningful output.
Exhibit 2: Large Language Models in practice
For an LLM to predict the next word in a sequence, it must be trained on vast amounts of data, specifically language-based data. This makes models particularly good at language-based tasks and therefore suitable to support mental health services, which are largely delivered verbally. More recently, AI models have been trained on multimodal data (such as images and audio) and to predict multimodal outputs, expanding their potential mental health use cases. For example, they could eventually be used to design images for imaginal exposures or provide audio-based guidance to clinicians through gentle suggestions.
Using LLMs in mental health settings
The data that LLM models are trained on can come from general sources (such as the internet) or from more tailored sources (such as psychology textbooks or therapy session transcripts). When models are trained on data from general sources, the data set is likely to include some information about mental health, but that information may not be evidence-based, may be biased, and may perpetuate mental health stigma, which affects the output that the model generates. As such, it is recommended that models be trained specifically to the contexts they will operate in, especially if they are being used in high-risk contexts such as healthcare and mental health.
There are multiple ways to prepare models to be trained to perform mental health tasks. One approach is to train the model using tailored data that is highly curated, evidence-based, representative, and specific to the mental health task the model will perform. Another approach is to take a model that was trained using data from general sources and fine-tune that model for a specific mental health use case using additional, tailored mental health data. Additional approaches—which can be used for models with data from either tailored or general sources—include giving the model “if-then” instructions for the types of outputs they should generate when presented with different mental health scenarios (in other words, prompt engineering) or having human mental health experts rate model output for quality and safety. The ratings would effectively teach the model how best to generate outputs (in other words, reinforcement learning with human feedback).
Models must be thoroughly tested before they are used with real people. This testing should involve a thorough evaluation to ensure that models are safe and unbiased prior to an initial launch. Additionally, safety and bias should be monitored on an ongoing basis.
When thoughtfully developed and responsibly deployed, AI could become a powerful ally in scaling mental health programs. In the next chapter, we explore how AI could enhance task sharing programs, highlighting illustrative AI use cases and real-world examples that also cover peer support and crisis counseling programs.