Project Overview
While enrolled at Carnegie Mellon, I participated in an independent study as a Research Assistant for Dr. Robert Kraut and Dr. Haiyi Zhu in the Human-Computer Interaction Institute. Together, on a cross-functional research team, we prototyped and designed AI personas that mimicked human behavior, speech patterns and symptoms of various mental health conditions.
To do this, we partnered with www.7cups.com, an online community of volunteer mental health paraprofessionals that provide emotional support to anonymous users through an online chat. The purpose of our developed AI personas was to provide virtual patients for these paraprofessionals to train and improve the quality of their mental healthcare provision. By training volunteers and mental health paraprofessionals, the ultimate goal is to expand the supply and accessibility of mental healthcare and support. This was an exploratory solution to the societal scarcity of affordable mental health support.
What I Learned:
Prompt Engineering is Part Art, Part Science
When building the AI patient personas, there were many iterations of trial and error. We borrowed from existing academic research on how to structure our prompts. We tried a variety of methods, ranging from chain prompting to multi-shot prompts. This scientific structure to the prompts was helpful, but part of our challenge was making the conversation seem realistic. In other words, we had to capture the speech patterns of various conditions and patient demographics in order to develop a realistic experience for the trainee.
In order to capture the realism and speech patterns needed for this training experience, it became necessary to blend details of our own life experiences and the experiences of past user conversations in order to come up with the personas. This included providing the personas realistic backstories, demographic attributes along with phrases and conversational instructions in order to make them seem real.
Developing a Persona Requires Technical Understanding
In order to develop strong virtual patient personas, we had to first build an operational understanding of large language models and natural language processing systems. This assisted us in being able to better predict how and why the attributes, demographic details and conversational parameters of our patient personas would be behave.
There is Potential for AI Innovation in Mental Healthcare
As part of our testing, we ran simulated conversations between virtual patients as well as AI-driven virtual therapists. Given the economic, social, technological and environmental factors of the present day, there are likely a wide range of product opportunity gaps that could be filled with a product similar to the virtual personas that we developed. These potential opportunities range from virtual companions, virtual therapists, virtual friends and even, as I explored in a later project, AI-driven augmented reality pets.